剖析虚幻渲染体系(12)- 移动端专题Part 1(UE移动端渲染分析)
阅读原文时间:2021年11月08日阅读:1

目录

12.1 本篇概述

前面的所有篇章都是基于PC端的延迟渲染管线阐述UE的渲染体系的,特别是剖析虚幻渲染体系(04)- 延迟渲染管线详尽地阐述了在PC端的延迟渲染管线的流程和步骤。

此篇只要针对UE的移动端的渲染管线进行阐述,最终还会对比移动和和PC端的渲染差异,以及特殊的优化措施。本篇主要阐述UE渲染体系的以下内容:

  • FMobileSceneRenderer的主要流程和步骤。
  • 移动端的前向和延迟渲染管线。
  • 移动端的光影和阴影。
  • 移动端和PC端的异同,以及涉及的特殊优化技巧。

特别要指出的是,本篇分析的UE源码升级到了4.27.1,需要同步看源码的同学注意更新了。

如果要在PC的UE编辑器开启移动端渲染管线,可以选择如下所示的菜单:

等待Shader编译完成,UE编辑器的视口内便是移动端的预览效果。

相比PC桌面平台,移动端在尺寸、电量、硬件性能等诸多方面都存在显著的差异,具体表现在:

  • 更小的尺寸。移动端的便携性就要求整机设备必须轻巧,可置于掌中或口袋内,所以整机只能限制在非常小的体积之内。

  • 有限的能量和功率。受限于电池存储技术,目前的主流锂电池普通在1万毫安,但移动设备的分辨率、画质却越来越高,为了满足足够长的续航和散热限制,必须严格控制移动设备的整机功率,通常在5w以内。

  • 散热方式受限。PC设备通常可以安装散热风扇、甚至水冷系统,而移动设备不具备这些主动散热方式,只能靠热传导散热。如果散热不当,CPU和GPU都会主动降频,以非常有限的性能运行,以免设备元器件因过热而损毁。

  • 有限的硬件性能。移动设备的各类元件(CPU、带宽、内存、GPU等)的性能都只是PC设备的数十分之一。

    2018年,主流PC设备(NV GV100-400-A1 Titan V)和主流移动设备(Samsung Exynos 9 8895)的性能对比图。移动设备的很多硬件性能只是PC设备的几十分之一,但分辨率却接近PC的一半,更加突显了移动设备的挑战和窘境。

    到了2020年,主流的移动设备性能如下所示:

  • 特殊的硬件架构。如CPU和GPU共享内存存储设备,被称为耦合式架构,还有GPU的TB(D)R架构,目的都是为了在低功耗内完成尽可能多的操作。

    PC设备的解耦硬件架构和移动设备的耦合式硬件架构对比图。

除此之外,不同于PC端的CPU和GPU纯粹地追求计算性能,衡量移动端的性能有三个指标:性能(Performance)、能量(Power)、面积(Area),俗称PPA。(下图)

衡量移动设备的三个基本参数:Performance、Area、Power,其中Compute Density(计算密度)涉及性能和面积,能耗比涉及性能和能力消耗,越大越好。

与移动设备一起崛起的还有XR设备,它是移动设备的一个重要发展分支。目前存在着各种不同大小、功能、应用场景的XR设备:

各种形式的XR设备。

随着近来元宇宙(Metaverse)的爆火,以及FaceBook改名为Meta,加之Apple、MicroSoft、NVidia、Google等科技巨头都在加紧布局面向未来的沉浸式体验,XR设备作为能够最接近元宇宙畅想的载体和入口,自然成为一条未来非常有潜力能够出现巨无霸的全新赛道。

12.2 UE移动端渲染特性

本章阐述一下UE4.27在移动端的渲染特性。

UE在移动端支持以下图形API:

Feature Level

说明

OpenGL ES 3.1

安卓系统的默认特性等级,可以在工程设置(Project Settings > Platforms > Android Material Quality - ES31)配置具体的材质参数。

Android Vulkan

可用于某些特定Android设备的高端渲染器,支持Vulkan 1.2 API,轻量级设计理念的Vulkan,多数情况下会比OpenGL更高效。

Metal 2.0

专用于iOS设备的特性等级。可以在Project Settings > Platforms > iOS Material Quality配置材质参数。

在目前的主流安卓设备,使用Vulkan能获得更好的性能,原因在于Vulkan轻量级的设计理念,使得UE等应用程序能够更加精准地执行优化。下面是Vulkan和OpenGL的对照表:

Vulkan

OpenGL

基于对象的状态,没有全局状态。

单一的全局状态机。

所有的状态概念都放置到命令缓冲区中。

状态被绑定到单个上下文。

可以多线程编码。

渲染操作只能被顺序执行。

可以精确、显式地操控GPU的内存和同步。

GPU的内存和同步细节通常被驱动程序隐藏起来。

驱动程序没有运行时错误检测,但存在针对开发人员的验证层。

广泛的运行时错误检测。

如果在Windows平台,UE编辑器也可以启动OpenGL、Vulkan、Metal的模拟器,以便在编辑器时预览效果,但可能跟实际的运行设备的画面有差异,不可完全依赖此功能。

开启Vulkan前需要在工程中配置一些参数,具体参看官方文档Android Vulkan Mobile Renderer

另外,UE早前几个版本就移除了windows下的OpenGL支持,虽然目前UE编辑器还存在OpenGL的模拟选项,但实际上底层是用D3D渲染的。

UE的Deferred Shading(延迟着色)是在4.26才加入的功能,使得开发者能够在移动端实现较复杂的光影效果,诸如高质量反射、多动态光照、贴花、高级光照特性。

上:前向渲染;下:延迟渲染。

如果要在移动端开启延迟渲染,需要在工程配置目录下的DefaultEngine.ini添加r.Mobile.ShadingPath=1字段,然后重启编辑器。

Ground Truth Ambient Occlusion (GTAO)是接近现实世界的环境遮挡技术,是阴影的一种补偿,能够遮蔽一部分非直接光照,从而获得良好的软阴影效果。

开启了GTAO的效果,注意机器人靠近墙面时,会在墙面留下具有渐变的软阴影效果。

为了开启GTAO,需要勾选以下所示的选项:

此外,GTAO依赖Mobile HDR的选项,为了在对应目标设备开启,还需要在[Platform]Scalability.ini的配置中添加r.Mobile.AmbientOcclusionQuality字段,并且值需要大于0,否则GTAO将被禁用。

值得注意的是,GTAO在Mali设备上存在性能问题,因为它们的最大Compute Shader线程数量少于1024个。

UE在移动端实现的光源特性有:

  • 线性空间的HDR光照。
  • 带方向的光照图(考虑了法线)。
  • 太阳(平行光)支持距离场阴影 + 解析的镜面高光。
  • IBL光照:每个物体采样了最近的一个反射捕捉器,而没有视差校正。
  • 动态物体能正确地接受光照,也可以投射阴影。

UE移动端支持的动态光源的类型、数量、阴影等信息如下:

光源类型

最大数量

阴影

描述

平行光

1

CSM

CSM默认是2级,最多支持4级。

点光源

4

不支持

点光源阴影需要立方体阴影图,而单Pass渲染立方体阴影(OnePassPointLightShadow)的技术需要GS(SM5才有)才支持。

聚光灯

4

支持

默认禁用,需要在工程中开启。

区域光

0

不支持

目前不支持动态区域光照效果。

动态的聚光灯需要在工程配置中显式开启:

在移动BasePass的像素着色器中,聚光灯阴影图与CSM共享相同的纹理采样器,并且聚光灯阴影和CSM使用相同的阴影图图集。CSM能够保证有足够的空间,而聚光灯将按阴影分辨率排序。

默认情况下,可见阴影的最大数量被限制为8个,但可以通过改变r.Mobile.MaxVisibleMovableSpotLightsShadow的值来改变上限值。聚光灯阴影的分辨率是基于屏幕大小和r.Shadow.TexelsPerPixelSpotlight

在前向渲染路径中,局部光源(点光源和聚光灯)的总数不能超过4个。

移动端还支持一种特殊的阴影模式,那就是调制阴影(Modulated Shadows),只能用于固定(Stationary)的平行光。开启了调制阴影的效果图如下:

调制阴影还支持改变阴影颜色和混合比例:

左:动态阴影;右:调制阴影。

移动端的阴影同样支持自阴影、阴影质量等级(r.shadowquality)、深度偏移等参数的设置。

此外,移动端默认使用了GGX的高光反射,如果想切换到传统的高光着色模型,可以在以下配置里修改:

UE针对移动端做了一个优化版的SSR,被称为Pixel Projected Reflection(PPR),也是复用屏幕空间像素的核心思想。

PPR效果图。

为了开启PPR效果,需要满足以下条件:

  • 开启MobileHDR选项。

  • r.Mobile.PixelProjectedReflectionQuality的值大于0。

  • 设置Project Settings > Mobile and set the Planar Reflection Mode成正确的模式:

    Planar Reflection Mode有3个选项:

    • Usual:平面反射Actor在所有平台上的作用都是相同的。
    • MobilePPR:平面反射Actor在PC/主机平台上正常工作,但在移动平台上使用PPR渲染。
    • MobilePPRExclusive:平面反射Actor将只用于移动平台上的PPR,为PC和Console项目留下了使用传统SSR的空间。

默认只有高端移动设备才会在[Project]Scalability.ini开启r.Mobile.PixelProjectedReflectionQuality

PC端的网格绘制管线已经支持了网格的自动化实例和合并绘制,这个特性可以极大提升渲染性能。4.27已经在移动端支持了这一特性。

若想开启,则需要打开工程配置目录下的DefaultEngine.ini,添加以下字段:

r.Mobile.SupportGPUScene=1
r.Mobile.UseGPUSceneTexture=1

重启编辑器,等待Shader编译完即可预览效果。

由于需要GPUSceneTexture支持,而Mali设备的Uniform Buffer最大只有64kb,以致无法支持足够大的空间,所以,Mali设备会使用纹理而非缓冲区来存储GPUScene数据。

但也存在一些限制:

  • 移动设备上的自动实例化主要有利于CPU密集型项目,而不是GPU密集型项目。虽然启用自动实例化不太可能会对GPU密集型的项目造成损害,但不太可能看到使用它带来的显著性能改进。

  • 如果一款游戏或应用需要大量内存,那么关闭r.Mobile.UseGPUSceneTexture并使用缓冲区可能会更有好处,因为它无法在Mali设备上正常运行。

    也可以针对Mali设备关闭r.Mobile.UseGPUSceneTexture,而其它GPU厂商的设备正常使用。

自动实例化的有效性很大程度上取决于项目的确切规范和定位,建议创建一个启用了自动实例化的构建,并对其进行概要分析,以确定是否会看到实质性的性能提升。

由于移动设备存在更慢的依赖纹理读取(dependent texture read)、有限的硬件特性、特殊的硬件架构、额外的渲染目标解析、有限的带宽等限制性因素,故而后处理在移动设备上执行起来会比较耗性能,有些极端情况会卡住渲染管线。

尽管如此,在某些画质要求高的游戏或应用,依然非常依赖后处理的强劲表现力,为高品质迈上几个台阶。UE不会限制开发者使用后处理。

为了开启后处理,必须先开启MobileHDR选项:

开启后处理之后,就可以在后处理体积(Post Process Volume)设置各种后处理效果。

在移动端可以支持的后处理有Mobile Tonemapper、Color Grading、Lens、Bloom、Dirt Mask、Auto Exposure、Lens Flares、Depth of Field等等。

为了获得更好的性能,官方给出的建议是在移动端只开启Bloom和TAA。

  • Reflection Capture Compression

移动端支持反射捕捉器组件(Reflection Capture Component)的压缩,可以减少Reflection Capture运行时的内存和带宽,提升渲染效率。需要在工程配置中开启:

开启之后,默认使用ETC2进行压缩。另外,也可以针对每个Reflection Capture Component进行调整:

  • 材质特性

移动平台上的材质(特性级别Open ES 3.1)使用与其他平台相同的基于节点的创建过程,并且绝大多数节点在移动端都支持。

移动平台支持的材质属性有:BaseColorRoughnessMetallicSpecularNormalEmissiveRefraction,但不支持Scene Color表达式、Tessellation输入、次表面散射着色模型。

移动平台支持的材质存在一些限制:

  • 由于硬件限制,只能使用16个纹理采样器。
  • 只有DefaultLit和Unlit着色模型可用。
  • 自定义UV应该用来避免依赖纹理读取(没有纹理uv的数学计算)。
  • 半透明和Masked材质是及其耗性能, 建议尽量使用不透明材质。
  • 深度淡出(Depth fade)可以在iOS平台的半透明材质中使用,但在硬件不支持从深度缓冲区获取数据的平台上,是不受支持的,将导致不可接受的性能成本。

材质属性面板存在一些针对移动端的特殊选项:

这些属性的说明如下:

  • Mobile Separate Translucency:是否在移动端开启单独的半透明渲染纹理。

  • Use Full Precision:是否使用全精度,如果是,可以减少带宽占用和能耗,提升性能,但可能会出现远处物体的瑕疵:

    左:全精度材质;右:半精度材质,远处的太阳出现了瑕疵。

  • Use Lightmap Directionality:是否开启光照图的方向性,若勾选,会考虑光照图的方向和像素法线,但会提升性能消耗。

  • Use Alpha to Coverage:是否为Masked材质开启MSAA抗锯齿,若勾选,会开启MSAA。

  • Fully Rough:是否完全粗糙,如果勾选,将极大提升此材质的渲染效率。

此外,移动端支持的网格类型有:

  • Skeletal Mesh
  • Static Mesh
  • Landscape
  • CPU particle sprites, particle meshes

除上述类型以外的其它都不被支持。其它限制还有:

  • 单个网格最多只能到65k,因为顶点索引只有16位。
  • 单个Skeletal Mesh的骨骼数量必须在75个以内,因为受硬件性能的限制。

12.3 FMobileSceneRenderer

FMobileSceneRenderer继承自FSceneRenderer,它负责移动端的场景渲染流程,而PC端是同样继承自FSceneRenderer的FDeferredShadingSceneRenderer。它们的继承关系图如下:

classDiagram-v2
FSceneRenderer <|-- FMobileSceneRenderer
FSceneRenderer <|-- FDeferredShadingSceneRenderer

前述多篇文章已经提及了FDeferredShadingSceneRenderer,它的渲染流程尤为复杂,包含了复杂的光影和渲染步骤。相比之下,FMobileSceneRenderer的逻辑和步骤会简单许多,下面是RenderDoc的截帧:

以上主要包含了InitViews、ShadowDepths、PrePass、BasePass、OcclusionTest、ShadowProjectionOnOpaque、Translucency、PostProcessing等步骤。其中这些步骤在PC端都是存在的,但实现过程可能会有所不同。见后续章节剖析。

移动端的场景渲染器的主流程也发生在FMobileSceneRenderer::Render中,代码和解析如下:

// Engine\Source\Runtime\Renderer\Private\MobileShadingRenderer.cpp

void FMobileSceneRenderer::Render(FRHICommandListImmediate& RHICmdList)
{
    // 更新图元场景信息。
    Scene->UpdateAllPrimitiveSceneInfos(RHICmdList);

    // 准备视图的渲染区域.
    PrepareViewRectsForRendering(RHICmdList);

    // 准备天空大气的数据
    if (ShouldRenderSkyAtmosphere(Scene, ViewFamily.EngineShowFlags))
    {
        for (int32 LightIndex = 0; LightIndex < NUM_ATMOSPHERE_LIGHTS; ++LightIndex)
        {
            if (Scene->AtmosphereLights[LightIndex])
            {
                PrepareSunLightProxy(*Scene->GetSkyAtmosphereSceneInfo(), LightIndex, *Scene->AtmosphereLights[LightIndex]);
            }
        }
    }
    else
    {
        Scene->ResetAtmosphereLightsProperties();
    }

    if(!ViewFamily.EngineShowFlags.Rendering)
    {
        return;
    }

    // 等待遮挡剔除测试.
    WaitOcclusionTests(RHICmdList);
    FRHICommandListExecutor::GetImmediateCommandList().PollOcclusionQueries();
    RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);

    // 初始化视图, 查找可见图元, 准备渲染所需的RT和缓冲区等数据.
    InitViews(RHICmdList);

    if (GRHINeedsExtraDeletionLatency || !GRHICommandList.Bypass())
    {
        QUICK_SCOPE_CYCLE_COUNTER(STAT_FMobileSceneRenderer_PostInitViewsFlushDel);

        // 可能会暂停遮挡查询,所以最好在等待时让RHI线程和GPU工作. 此外,当执行RHI线程时,这是唯一将处理挂起删除的位置.
        FRHICommandListExecutor::GetImmediateCommandList().PollOcclusionQueries();
        FRHICommandListExecutor::GetImmediateCommandList().ImmediateFlush(EImmediateFlushType::FlushRHIThreadFlushResources);
    }

    GEngine->GetPreRenderDelegate().Broadcast();

    // 在渲染开始前提交全局动态缓冲.
    DynamicIndexBuffer.Commit();
    DynamicVertexBuffer.Commit();
    DynamicReadBuffer.Commit();
    RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);

    RHICmdList.SetCurrentStat(GET_STATID(STAT_CLMM_SceneSim));

    if (ViewFamily.bLateLatchingEnabled)
    {
        BeginLateLatching(RHICmdList);
    }

    FSceneRenderTargets& SceneContext = FSceneRenderTargets::Get(RHICmdList);

    // 处理虚拟纹理
    if (bUseVirtualTexturing)
    {
        SCOPED_GPU_STAT(RHICmdList, VirtualTextureUpdate);

        FVirtualTextureSystem::Get().Update(RHICmdList, FeatureLevel, Scene);

        // Clear virtual texture feedback to default value
        FUnorderedAccessViewRHIRef FeedbackUAV = SceneContext.GetVirtualTextureFeedbackUAV();
        RHICmdList.Transition(FRHITransitionInfo(FeedbackUAV, ERHIAccess::SRVMask, ERHIAccess::UAVMask));
        RHICmdList.ClearUAVUint(FeedbackUAV, FUintVector4(~0u, ~0u, ~0u, ~0u));
        RHICmdList.Transition(FRHITransitionInfo(FeedbackUAV, ERHIAccess::UAVMask, ERHIAccess::UAVMask));
        RHICmdList.BeginUAVOverlap(FeedbackUAV);
    }

    // 已排序的光源信息.
    FSortedLightSetSceneInfo SortedLightSet;

    // 延迟渲染.
    if (bDeferredShading)
    {
        // 收集和排序光源.
        GatherAndSortLights(SortedLightSet);

        int32 NumReflectionCaptures = Views[0].NumBoxReflectionCaptures + Views[0].NumSphereReflectionCaptures;
        bool bCullLightsToGrid = (NumReflectionCaptures > 0 || GMobileUseClusteredDeferredShading != 0);
        FRDGBuilder GraphBuilder(RHICmdList);
        // 计算光源格子.
        ComputeLightGrid(GraphBuilder, bCullLightsToGrid, SortedLightSet);
        GraphBuilder.Execute();
    }

    // 生成天空/大气LUT.
    const bool bShouldRenderSkyAtmosphere = ShouldRenderSkyAtmosphere(Scene, ViewFamily.EngineShowFlags);
    if (bShouldRenderSkyAtmosphere)
    {
        FRDGBuilder GraphBuilder(RHICmdList);
        RenderSkyAtmosphereLookUpTables(GraphBuilder);
        GraphBuilder.Execute();
    }

    // 通知特效系统场景准备渲染.
    if (FXSystem && ViewFamily.EngineShowFlags.Particles)
    {
        FXSystem->PreRender(RHICmdList, NULL, !Views[0].bIsPlanarReflection);
        if (FGPUSortManager* GPUSortManager = FXSystem->GetGPUSortManager())
        {
            GPUSortManager->OnPreRender(RHICmdList);
        }
    }
    // 轮询遮挡剔除请求.
    FRHICommandListExecutor::GetImmediateCommandList().PollOcclusionQueries();
    RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);

    RHICmdList.SetCurrentStat(GET_STATID(STAT_CLMM_Shadows));

    // 渲染阴影.
    RenderShadowDepthMaps(RHICmdList);
    FRHICommandListExecutor::GetImmediateCommandList().PollOcclusionQueries();
    RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);

    // 收集视图列表.
    TArray<const FViewInfo*> ViewList;
    for (int32 ViewIndex = 0; ViewIndex < Views.Num(); ViewIndex++)
    {
        ViewList.Add(&Views[ViewIndex]);
    }

    // 渲染自定义深度.
    if (bShouldRenderCustomDepth)
    {
        FRDGBuilder GraphBuilder(RHICmdList);
        FSceneTextureShaderParameters SceneTextures = CreateSceneTextureShaderParameters(GraphBuilder, Views[0].GetFeatureLevel(), ESceneTextureSetupMode::None);
        RenderCustomDepthPass(GraphBuilder, SceneTextures);
        GraphBuilder.Execute();
    }

    // 渲染深度PrePass.
    if (bIsFullPrepassEnabled)
    {
        // SDF和AO需要完整的PrePass深度.

        FRHIRenderPassInfo DepthPrePassRenderPassInfo(
            SceneContext.GetSceneDepthSurface(),
            EDepthStencilTargetActions::ClearDepthStencil_StoreDepthStencil);

        DepthPrePassRenderPassInfo.NumOcclusionQueries = ComputeNumOcclusionQueriesToBatch();
        DepthPrePassRenderPassInfo.bOcclusionQueries = DepthPrePassRenderPassInfo.NumOcclusionQueries != 0;

        RHICmdList.BeginRenderPass(DepthPrePassRenderPassInfo, TEXT("DepthPrepass"));

        RHICmdList.SetCurrentStat(GET_STATID(STAT_CLM_MobilePrePass));

        // 渲染完整的深度PrePass.
        RenderPrePass(RHICmdList);

        // 提交遮挡剔除.
        RHICmdList.SetCurrentStat(GET_STATID(STAT_CLMM_Occlusion));
        RenderOcclusion(RHICmdList);

        RHICmdList.EndRenderPass();

        // SDF阴影
        if (bRequiresDistanceFieldShadowingPass)
        {
            CSV_SCOPED_TIMING_STAT_EXCLUSIVE(RenderSDFShadowing);
            RenderSDFShadowing(RHICmdList);
        }

        // HZB.
        if (bShouldRenderHZB)
        {
            RenderHZB(RHICmdList, SceneContext.SceneDepthZ);
        }

        // AO.
        if (bRequiresAmbientOcclusionPass)
        {
            RenderAmbientOcclusion(RHICmdList, SceneContext.SceneDepthZ);
        }
    }

    FRHITexture* SceneColor = nullptr;

    // 延迟渲染.
    if (bDeferredShading)
    {
        SceneColor = RenderDeferred(RHICmdList, ViewList, SortedLightSet);
    }
    // 前向渲染.
    else
    {
        SceneColor = RenderForward(RHICmdList, ViewList);
    }

    // 渲染速度缓冲.
    if (bShouldRenderVelocities)
    {
        FRDGBuilder GraphBuilder(RHICmdList);

        FRDGTextureMSAA SceneDepthTexture = RegisterExternalTextureMSAA(GraphBuilder, SceneContext.SceneDepthZ);
        FRDGTextureRef VelocityTexture = TryRegisterExternalTexture(GraphBuilder, SceneContext.SceneVelocity);

        if (VelocityTexture != nullptr)
        {
            AddClearRenderTargetPass(GraphBuilder, VelocityTexture);
        }

        // 渲染可移动物体的速度缓冲.
        AddSetCurrentStatPass(GraphBuilder, GET_STATID(STAT_CLMM_Velocity));
        RenderVelocities(GraphBuilder, SceneDepthTexture.Resolve, VelocityTexture, FSceneTextureShaderParameters(), EVelocityPass::Opaque, false);
        AddSetCurrentStatPass(GraphBuilder, GET_STATID(STAT_CLMM_AfterVelocity));

        // 渲染透明物体的速度缓冲.
        AddSetCurrentStatPass(GraphBuilder, GET_STATID(STAT_CLMM_TranslucentVelocity));
        RenderVelocities(GraphBuilder, SceneDepthTexture.Resolve, VelocityTexture, GetSceneTextureShaderParameters(CreateMobileSceneTextureUniformBuffer(GraphBuilder, EMobileSceneTextureSetupMode::SceneColor)), EVelocityPass::Translucent, false);

        GraphBuilder.Execute();
    }

    // 处理场景渲染后的逻辑.
    {
        FRendererModule& RendererModule = static_cast<FRendererModule&>(GetRendererModule());
        FRDGBuilder GraphBuilder(RHICmdList);
        RendererModule.RenderPostOpaqueExtensions(GraphBuilder, Views, SceneContext);

        if (FXSystem && Views.IsValidIndex(0))
        {
            AddUntrackedAccessPass(GraphBuilder, [this](FRHICommandListImmediate& RHICmdList)
            {
                check(RHICmdList.IsOutsideRenderPass());

                FXSystem->PostRenderOpaque(
                    RHICmdList,
                    Views[0].ViewUniformBuffer,
                    nullptr,
                    nullptr,
                    Views[0].AllowGPUParticleUpdate()
                );
                if (FGPUSortManager* GPUSortManager = FXSystem->GetGPUSortManager())
                {
                    GPUSortManager->OnPostRenderOpaque(RHICmdList);
                }
            });
        }
        GraphBuilder.Execute();
    }

    // 刷新/提交命令缓冲.
    if (bSubmitOffscreenRendering)
    {
        RHICmdList.SubmitCommandsHint();
        RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);
    }

    // 转换场景颜色成SRV, 以供后续步骤读取.
    if (!bGammaSpace || bRenderToSceneColor)
    {
        RHICmdList.Transition(FRHITransitionInfo(SceneColor, ERHIAccess::Unknown, ERHIAccess::SRVMask));
    }

    if (bDeferredShading)
    {
        // 释放场景渲染目标上的原始引用.
        SceneContext.AdjustGBufferRefCount(RHICmdList, -1);
    }

    RHICmdList.SetCurrentStat(GET_STATID(STAT_CLMM_Post));

    // 处理虚拟纹理.
    if (bUseVirtualTexturing)
    {
        SCOPED_GPU_STAT(RHICmdList, VirtualTextureUpdate);

        // No pass after this should make VT page requests
        RHICmdList.EndUAVOverlap(SceneContext.VirtualTextureFeedbackUAV);
        RHICmdList.Transition(FRHITransitionInfo(SceneContext.VirtualTextureFeedbackUAV, ERHIAccess::UAVMask, ERHIAccess::SRVMask));

        TArray<FIntRect, TInlineAllocator<4>> ViewRects;
        ViewRects.AddUninitialized(Views.Num());
        for (int32 ViewIndex = 0; ViewIndex < Views.Num(); ++ViewIndex)
        {
            ViewRects[ViewIndex] = Views[ViewIndex].ViewRect;
        }

        FVirtualTextureFeedbackBufferDesc Desc;
        Desc.Init2D(SceneContext.GetBufferSizeXY(), ViewRects, SceneContext.GetVirtualTextureFeedbackScale());

        SubmitVirtualTextureFeedbackBuffer(RHICmdList, SceneContext.VirtualTextureFeedback, Desc);
    }

    FMemMark Mark(FMemStack::Get());
    FRDGBuilder GraphBuilder(RHICmdList);

    FRDGTextureRef ViewFamilyTexture = TryCreateViewFamilyTexture(GraphBuilder, ViewFamily);

    // 解析场景
    if (ViewFamily.bResolveScene)
    {
        if (!bGammaSpace || bRenderToSceneColor)
        {
            // 完成每个视图的渲染或完整的立体声缓冲区(如果启用)
            {
                RDG_EVENT_SCOPE(GraphBuilder, "PostProcessing");
                SCOPE_CYCLE_COUNTER(STAT_FinishRenderViewTargetTime);

                TArray<TRDGUniformBufferRef<FMobileSceneTextureUniformParameters>, TInlineAllocator<1, SceneRenderingAllocator>> MobileSceneTexturesPerView;
                MobileSceneTexturesPerView.SetNumZeroed(Views.Num());

                const auto SetupMobileSceneTexturesPerView = [&]()
                {
                    for (int32 ViewIndex = 0; ViewIndex < Views.Num(); ++ViewIndex)
                    {
                        EMobileSceneTextureSetupMode SetupMode = EMobileSceneTextureSetupMode::SceneColor;
                        if (Views[ViewIndex].bCustomDepthStencilValid)
                        {
                            SetupMode |= EMobileSceneTextureSetupMode::CustomDepth;
                        }

                        if (bShouldRenderVelocities)
                        {
                            SetupMode |= EMobileSceneTextureSetupMode::SceneVelocity;
                        }

                        MobileSceneTexturesPerView[ViewIndex] = CreateMobileSceneTextureUniformBuffer(GraphBuilder, SetupMode);
                    }
                };

                SetupMobileSceneTexturesPerView();

                FMobilePostProcessingInputs PostProcessingInputs;
                PostProcessingInputs.ViewFamilyTexture = ViewFamilyTexture;

                // 渲染后处理效果.
                for (int32 ViewIndex = 0; ViewIndex < Views.Num(); ViewIndex++)
                {
                    RDG_EVENT_SCOPE_CONDITIONAL(GraphBuilder, Views.Num() > 1, "View%d", ViewIndex);
                    PostProcessingInputs.SceneTextures = MobileSceneTexturesPerView[ViewIndex];
                    AddMobilePostProcessingPasses(GraphBuilder, Views[ViewIndex], PostProcessingInputs, NumMSAASamples > 1);
                }
            }
        }
    }

    GEngine->GetPostRenderDelegate().Broadcast();

    RHICmdList.SetCurrentStat(GET_STATID(STAT_CLMM_SceneEnd));

    if (bShouldRenderVelocities)
    {
        SceneContext.SceneVelocity.SafeRelease();
    }

    if (ViewFamily.bLateLatchingEnabled)
    {
        EndLateLatching(RHICmdList, Views[0]);
    }

    RenderFinish(GraphBuilder, ViewFamilyTexture);
    GraphBuilder.Execute();

    // 轮询遮挡剔除请求.
    FRHICommandListExecutor::GetImmediateCommandList().PollOcclusionQueries();
    FRHICommandListExecutor::GetImmediateCommandList().ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);
}

看过剖析虚幻渲染体系(04)- 延迟渲染管线篇章的同学应该都知道,移动端的场景渲染过程精简了很多很多步骤,相当于是PC端场景渲染器的一个子集。当然,为了适应移动端特有的GPU硬件架构,移动端的场景渲染也有区别于PC端的地方。后面会详细剖析。移动端场景的主要步骤和流程如下所示:

stateDiagram-v2
state bDeferredShading <>
state bIsFullPrepassEnabled <>

[*] --> UpdateAllPrimitiveSceneInfos
UpdateAllPrimitiveSceneInfos --> PrepareViewRectsForRendering
PrepareViewRectsForRendering --> InitViews
InitViews --> bDeferredShading
bDeferredShading --> RenderSkyAtmosphereLookUpTables* : No
bDeferredShading --> GatherAndSortLights: Yes
GatherAndSortLights --> ComputeLightGrid
ComputeLightGrid --> RenderSkyAtmosphereLookUpTables*

RenderSkyAtmosphereLookUpTables* --> RenderShadowDepthMaps
RenderShadowDepthMaps --> RenderCustomDepthPass*
RenderCustomDepthPass* --> bIsFullPrepassEnabled

bIsFullPrepassEnabled --> bDeferredShading2: No
bIsFullPrepassEnabled --> RenderPrePass:Yes
RenderPrePass --> RenderOcclusion
RenderOcclusion --> RenderSDFShadowing*
RenderSDFShadowing* --> RenderHZB*
RenderHZB* --> RenderAmbientOcclusion*

RenderAmbientOcclusion* --> bDeferredShading2

bDeferredShading2-->RenderForward:No
bDeferredShading2-->RenderDeferred:Yes

RenderDeferred --> RenderVelocities*
RenderForward --> RenderVelocities*

RenderVelocities* --> AddMobilePostProcessingPasses*
AddMobilePostProcessingPasses*-->RenderFinish
RenderFinish --> [*]

关于上面的流程图,有以下几点需要加以说明:

  • 流程图节点bDeferredShadingbDeferredShading2是同一个变量,这里区分开主要是为了防止mermaid语法绘图错误。
  • 带*的节点是有条件的,非必然执行的步骤。

UE4.26便加入了移动端的延迟渲染管线,所以上述代码中有前向渲染分支RenderForward和延迟渲染分支RenderDeferred,它们返回的都是渲染结果SceneColor。

移动端也支持了图元GPU场景、SDF阴影、AO、天空大气、虚拟纹理、遮挡剔除等渲染特性。

自UE4.26开始,渲染体系广泛地使用了RDG系统,移动端的场景渲染器也不例外。上述代码中总共声明了数个FRDGBuilder实例,用于计算光源格子,以及渲染天空大气LUT、自定义深度、速度缓冲、渲染后置事件、后处理等,它们都是相对独立的功能模块或渲染阶段。

RenderForward在移动端场景渲染器中负责前向渲染的分支,它的代码和解析如下:

FRHITexture* FMobileSceneRenderer::RenderForward(FRHICommandListImmediate& RHICmdList, const TArrayView<const FViewInfo*> ViewList)
{
    const FViewInfo& View = *ViewList[0];
    FSceneRenderTargets& SceneContext = FSceneRenderTargets::Get(RHICmdList);

    FRHITexture* SceneColor = nullptr;
    FRHITexture* SceneColorResolve = nullptr;
    FRHITexture* SceneDepth = nullptr;
    ERenderTargetActions ColorTargetAction = ERenderTargetActions::Clear_Store;
    EDepthStencilTargetActions DepthTargetAction = EDepthStencilTargetActions::ClearDepthStencil_DontStoreDepthStencil;

    // 是否启用移动端MSAA.
    bool bMobileMSAA = NumMSAASamples > 1 && SceneContext.GetSceneColorSurface()->GetNumSamples() > 1;

    // 是否启用移动端多试图模式.
    static const auto CVarMobileMultiView = IConsoleManager::Get().FindTConsoleVariableDataInt(TEXT("vr.MobileMultiView"));
    const bool bIsMultiViewApplication = (CVarMobileMultiView && CVarMobileMultiView->GetValueOnAnyThread() != 0);

    // gamma空间的渲染分支.
    if (bGammaSpace && !bRenderToSceneColor)
    {
        // 如果开启MSAA, 则从SceneContext获取渲染纹理(包含场景颜色和解析纹理)
        if (bMobileMSAA)
        {
            SceneColor = SceneContext.GetSceneColorSurface();
            SceneColorResolve = ViewFamily.RenderTarget->GetRenderTargetTexture();
            ColorTargetAction = ERenderTargetActions::Clear_Resolve;
            RHICmdList.Transition(FRHITransitionInfo(SceneColorResolve, ERHIAccess::Unknown, ERHIAccess::RTV | ERHIAccess::ResolveDst));
        }
        // 非MSAA,从视图家族获取渲染纹理.
        else
        {
            SceneColor = ViewFamily.RenderTarget->GetRenderTargetTexture();
            RHICmdList.Transition(FRHITransitionInfo(SceneColor, ERHIAccess::Unknown, ERHIAccess::RTV));
        }
        SceneDepth = SceneContext.GetSceneDepthSurface();
    }
    // 线性空间或渲染到场景纹理.
    else
    {
        SceneColor = SceneContext.GetSceneColorSurface();
        if (bMobileMSAA)
        {
            SceneColorResolve = SceneContext.GetSceneColorTexture();
            ColorTargetAction = ERenderTargetActions::Clear_Resolve;
            RHICmdList.Transition(FRHITransitionInfo(SceneColorResolve, ERHIAccess::Unknown, ERHIAccess::RTV | ERHIAccess::ResolveDst));
        }
        else
        {
            SceneColorResolve = nullptr;
            ColorTargetAction = ERenderTargetActions::Clear_Store;
        }

        SceneDepth = SceneContext.GetSceneDepthSurface();

        if (bRequiresMultiPass)
        {
            // store targets after opaque so translucency render pass can be restarted
            ColorTargetAction = ERenderTargetActions::Clear_Store;
            DepthTargetAction = EDepthStencilTargetActions::ClearDepthStencil_StoreDepthStencil;
        }

        if (bKeepDepthContent)
        {
            // store depth if post-processing/capture needs it
            DepthTargetAction = EDepthStencilTargetActions::ClearDepthStencil_StoreDepthStencil;
        }
    }

    // prepass的深度纹理状态.
    if (bIsFullPrepassEnabled)
    {
        ERenderTargetActions DepthTarget = MakeRenderTargetActions(ERenderTargetLoadAction::ELoad, GetStoreAction(GetDepthActions(DepthTargetAction)));
        ERenderTargetActions StencilTarget = MakeRenderTargetActions(ERenderTargetLoadAction::ELoad, GetStoreAction(GetStencilActions(DepthTargetAction)));
        DepthTargetAction = MakeDepthStencilTargetActions(DepthTarget, StencilTarget);
    }

    FRHITexture* ShadingRateTexture = nullptr;

    if (!View.bIsSceneCapture && !View.bIsReflectionCapture)
    {
        TRefCountPtr<IPooledRenderTarget> ShadingRateTarget = GVRSImageManager.GetMobileVariableRateShadingImage(ViewFamily);
        if (ShadingRateTarget.IsValid())
        {
            ShadingRateTexture = ShadingRateTarget->GetRenderTargetItem().ShaderResourceTexture;
        }
    }

    // 场景颜色渲染Pass信息.
    FRHIRenderPassInfo SceneColorRenderPassInfo(
        SceneColor,
        ColorTargetAction,
        SceneColorResolve,
        SceneDepth,
        DepthTargetAction,
        nullptr, // we never resolve scene depth on mobile
        ShadingRateTexture,
        VRSRB_Sum,
        FExclusiveDepthStencil::DepthWrite_StencilWrite
    );
    SceneColorRenderPassInfo.SubpassHint = ESubpassHint::DepthReadSubpass;
    if (!bIsFullPrepassEnabled)
    {
        SceneColorRenderPassInfo.NumOcclusionQueries = ComputeNumOcclusionQueriesToBatch();
        SceneColorRenderPassInfo.bOcclusionQueries = SceneColorRenderPassInfo.NumOcclusionQueries != 0;
    }
    // 如果场景颜色不是多视图,但应用程序是,需要渲染为单视图的多视图给着色器.
    SceneColorRenderPassInfo.MultiViewCount = View.bIsMobileMultiViewEnabled ? 2 : (bIsMultiViewApplication ? 1 : 0);

    // 开始渲染场景颜色.
    RHICmdList.BeginRenderPass(SceneColorRenderPassInfo, TEXT("SceneColorRendering"));

    if (GIsEditor && !View.bIsSceneCapture)
    {
        DrawClearQuad(RHICmdList, Views[0].BackgroundColor);
    }

    if (!bIsFullPrepassEnabled)
    {
        RHICmdList.SetCurrentStat(GET_STATID(STAT_CLM_MobilePrePass));
        // 渲染深度pre-pass
        RenderPrePass(RHICmdList);
    }

    RHICmdList.SetCurrentStat(GET_STATID(STAT_CLMM_Opaque));
    // 渲染BasePass: 不透明和masked物体.
    RenderMobileBasePass(RHICmdList, ViewList);
    RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);

    //渲染调试模式.
#if !(UE_BUILD_SHIPPING || UE_BUILD_TEST)
    if (ViewFamily.UseDebugViewPS())
    {
        // Here we use the base pass depth result to get z culling for opaque and masque.
        // The color needs to be cleared at this point since shader complexity renders in additive.
        DrawClearQuad(RHICmdList, FLinearColor::Black);
        RenderMobileDebugView(RHICmdList, ViewList);
        RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);
    }
#endif // !(UE_BUILD_SHIPPING || UE_BUILD_TEST)

    const bool bAdrenoOcclusionMode = CVarMobileAdrenoOcclusionMode.GetValueOnRenderThread() != 0;
    if (!bIsFullPrepassEnabled)
    {
        // 遮挡剔除
        if (!bAdrenoOcclusionMode)
        {
            // 提交遮挡剔除
            RHICmdList.SetCurrentStat(GET_STATID(STAT_CLMM_Occlusion));
            RenderOcclusion(RHICmdList);
            RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);
        }
    }

    // 后置事件, 处理插件渲染.
    {
        CSV_SCOPED_TIMING_STAT_EXCLUSIVE(ViewExtensionPostRenderBasePass);
        QUICK_SCOPE_CYCLE_COUNTER(STAT_FMobileSceneRenderer_ViewExtensionPostRenderBasePass);
        for (int32 ViewExt = 0; ViewExt < ViewFamily.ViewExtensions.Num(); ++ViewExt)
        {
            for (int32 ViewIndex = 0; ViewIndex < ViewFamily.Views.Num(); ++ViewIndex)
            {
                ViewFamily.ViewExtensions[ViewExt]->PostRenderBasePass_RenderThread(RHICmdList, Views[ViewIndex]);
            }
        }
    }

    // 如果需要渲染透明物体或像素投影的反射, 则需要拆分pass.
    if (bRequiresMultiPass || bRequiresPixelProjectedPlanarRelfectionPass)
    {
        RHICmdList.EndRenderPass();
    }

    RHICmdList.SetCurrentStat(GET_STATID(STAT_CLMM_Translucency));

    // 如果需要, 则重新开启透明渲染通道.
    if (bRequiresMultiPass || bRequiresPixelProjectedPlanarRelfectionPass)
    {
        check(RHICmdList.IsOutsideRenderPass());

        // 如果当前硬件不支持读写相同的深度缓冲区,则复制场景深度.
        ConditionalResolveSceneDepth(RHICmdList, View);

        if (bRequiresPixelProjectedPlanarRelfectionPass)
        {
            const FPlanarReflectionSceneProxy* PlanarReflectionSceneProxy = Scene ? Scene->GetForwardPassGlobalPlanarReflection() : nullptr;
            RenderPixelProjectedReflection(RHICmdList, SceneContext, PlanarReflectionSceneProxy);

            FRHITransitionInfo TranslucentRenderPassTransitions[] = {
            FRHITransitionInfo(SceneColor, ERHIAccess::SRVMask, ERHIAccess::RTV),
            FRHITransitionInfo(SceneDepth, ERHIAccess::SRVMask, ERHIAccess::DSVWrite)
            };
            RHICmdList.Transition(MakeArrayView(TranslucentRenderPassTransitions, UE_ARRAY_COUNT(TranslucentRenderPassTransitions)));
        }

        DepthTargetAction = EDepthStencilTargetActions::LoadDepthStencil_DontStoreDepthStencil;
        FExclusiveDepthStencil::Type ExclusiveDepthStencil = FExclusiveDepthStencil::DepthRead_StencilRead;
        if (bModulatedShadowsInUse)
        {
            ExclusiveDepthStencil = FExclusiveDepthStencil::DepthRead_StencilWrite;
        }

        // 用于移动端像素投影反射的不透明网格必须将深度写入深度RT, 因为只渲染一次网格(如果质量水平低于或等于BestPerformance).
        if (IsMobilePixelProjectedReflectionEnabled(View.GetShaderPlatform())
            && GetMobilePixelProjectedReflectionQuality() == EMobilePixelProjectedReflectionQuality::BestPerformance)
        {
            ExclusiveDepthStencil = FExclusiveDepthStencil::DepthWrite_StencilWrite;
        }

        if (bKeepDepthContent && !bMobileMSAA)
        {
            DepthTargetAction = EDepthStencilTargetActions::LoadDepthStencil_StoreDepthStencil;
        }

#if PLATFORM_HOLOLENS
        if (bShouldRenderDepthToTranslucency)
        {
            ExclusiveDepthStencil = FExclusiveDepthStencil::DepthWrite_StencilWrite;
        }
#endif

        // 透明物体渲染Pass.
        FRHIRenderPassInfo TranslucentRenderPassInfo(
            SceneColor,
            SceneColorResolve ? ERenderTargetActions::Load_Resolve : ERenderTargetActions::Load_Store,
            SceneColorResolve,
            SceneDepth,
            DepthTargetAction,
            nullptr,
            ShadingRateTexture,
            VRSRB_Sum,
            ExclusiveDepthStencil
        );
        TranslucentRenderPassInfo.NumOcclusionQueries = 0;
        TranslucentRenderPassInfo.bOcclusionQueries = false;
        TranslucentRenderPassInfo.SubpassHint = ESubpassHint::DepthReadSubpass;

        // 开始渲染半透明物体.
        RHICmdList.BeginRenderPass(TranslucentRenderPassInfo, TEXT("SceneColorTranslucencyRendering"));
    }

    // 场景深度是只读的,可以获取.
    RHICmdList.NextSubpass();

    if (!View.bIsPlanarReflection)
    {
        // 渲染贴花.
        if (ViewFamily.EngineShowFlags.Decals)
        {
            CSV_SCOPED_TIMING_STAT_EXCLUSIVE(RenderDecals);
            RenderDecals(RHICmdList);
        }

        // 渲染调制阴影投射.
        if (ViewFamily.EngineShowFlags.DynamicShadows)
        {
            CSV_SCOPED_TIMING_STAT_EXCLUSIVE(RenderShadowProjections);
            RenderModulatedShadowProjections(RHICmdList);
        }
    }

    // 绘制半透明.
    if (ViewFamily.EngineShowFlags.Translucency)
    {
        CSV_SCOPED_TIMING_STAT_EXCLUSIVE(RenderTranslucency);
        SCOPE_CYCLE_COUNTER(STAT_TranslucencyDrawTime);

        RenderTranslucency(RHICmdList, ViewList);

        FRHICommandListExecutor::GetImmediateCommandList().PollOcclusionQueries();
        RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);
    }

    if (!bIsFullPrepassEnabled)
    {
        // Adreno遮挡剔除模式.
        if (bAdrenoOcclusionMode)
        {
            RHICmdList.SetCurrentStat(GET_STATID(STAT_CLMM_Occlusion));
            // flush
            RHICmdList.SubmitCommandsHint();
            bSubmitOffscreenRendering = false; // submit once
            // Issue occlusion queries
            RenderOcclusion(RHICmdList);
            RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);
        }
    }

    // 在MSAA被解析前预计算色调映射(只在iOS有效)
    if (!bGammaSpace)
    {
        PreTonemapMSAA(RHICmdList);
    }

    // 结束场景颜色渲染.
    RHICmdList.EndRenderPass();

    // 优化返回场景颜色的解析纹理(开启了MSAA才有).
    return SceneColorResolve ? SceneColorResolve : SceneColor;
}

移动端前向渲染主要步骤跟PC端类似,依次渲染PrePass、BasePass、特殊渲染(贴花、AO、遮挡剔除等)、半透明物体。它们的流程图如下:

stateDiagram-v2
[*] --> DrawClearQuad*

DrawClearQuad* --> RenderPrePass*
RenderPrePass* --> RenderMobileBasePass
RenderMobileBasePass --> RenderOcclusion*

RenderOcclusion* --> RenderDecals*
RenderDecals* --> RenderModulatedShadowProjections*
RenderModulatedShadowProjections* --> RenderTranslucency*
RenderTranslucency* --> PreTonemapMSAA*

PreTonemapMSAA* --> [*]

其中遮挡剔除和GPU厂商相关,比如高通Adreno系列GPU芯片要求在Flush渲染指令和Switch FBO之间:

Render Opaque -> Render Translucent -> Flush -> Render Queries -> Switch FBO

那么UE也遵循了Adreno系列芯片的特殊要求,对其的遮挡剔除做了特殊的处理。

Adreno系列芯片支持TBDR架构的Bin和普通的Direct两种混合模式的渲染,会在遮挡查询时自动切换到Direct模式,以降低遮挡查询的开销。如果不在Flush渲染指令和Switch FBO之间提交查询,会卡住整个渲染管线,引发渲染性能下降。

MSAA由于天然硬件支持且效果和效率达到很好的平衡,是UE在移动端前向渲染的首选抗锯齿。因此,上述代码中出现了不少处理MSAA的逻辑,包含颜色和深度纹理及其资源状态。如果开启了MSAA,默认情况下是在RHICmdList.EndRenderPass()解析场景颜色(同时将芯片分块上的数据写回到系统显存中),由此获得抗锯齿的纹理。移动端的MSAA默认不开启,但可在以下界面中设置:

前向渲染支持Gamma空间和HDR(线性空间)两种颜色空间模式。如果是线性空间,则渲染后期需要色调映射等步骤。默认是HDR,可在项目配置中更改:

上述代码的bRequiresMultiPass标明了是否需要专用的渲染Pass绘制半透明物体,决定它的值由以下代码完成:

// Engine\Source\Runtime\Renderer\Private\MobileShadingRenderer.cpp

bool FMobileSceneRenderer::RequiresMultiPass(FRHICommandListImmediate& RHICmdList, const FViewInfo& View) const
{
    // Vulkan uses subpasses
    if (IsVulkanPlatform(ShaderPlatform))
    {
        return false;
    }

    // All iOS support frame_buffer_fetch
    if (IsMetalMobilePlatform(ShaderPlatform))
    {
        return false;
    }

    if (IsMobileDeferredShadingEnabled(ShaderPlatform))
    {
        // TODO: add GL support
        return true;
    }

    // Some Androids support frame_buffer_fetch
    if (IsAndroidOpenGLESPlatform(ShaderPlatform) && (GSupportsShaderFramebufferFetch || GSupportsShaderDepthStencilFetch))
    {
        return false;
    }

    // Always render reflection capture in single pass
    if (View.bIsPlanarReflection || View.bIsSceneCapture)
    {
        return false;
    }

    // Always render LDR in single pass
    if (!IsMobileHDR())
    {
        return false;
    }

    // MSAA depth can't be sampled or resolved, unless we are on PC (no vulkan)
    if (NumMSAASamples > 1 && !IsSimulatedPlatform(ShaderPlatform))
    {
        return false;
    }

    return true;
}

与之类似但意义不同的是bIsMultiViewApplication和bIsMobileMultiViewEnabled标记,标明是否开启多视图渲染以及多视图的个数。只用于VR,由控制台变量vr.MobileMultiView及图形API等因素决定。MultiView用于XR,用于优化渲染两次的情形,它存在Basic和Advanced两种模式:

用于优化VR等渲染的MultiView对比图。上:未采用MultiView模式的渲染,两个眼睛各自提交绘制指令;中:基础MultiView模式,复用提交指令,在GPU层复制多一份Command List;下:高级MultiView模式,可以复用DC、Command List、几何信息。

bKeepDepthContent标明是否要保留深度内容,决定它的代码:

bKeepDepthContent =
    bRequiresMultiPass ||
    bForceDepthResolve ||
    bRequiresPixelProjectedPlanarRelfectionPass ||
    bSeparateTranslucencyActive ||
    Views[0].bIsReflectionCapture ||
    (bDeferredShading && bPostProcessUsesSceneDepth) ||
    bShouldRenderVelocities ||
    bIsFullPrepassEnabled;

// 带MSAA的深度从不保留.
bKeepDepthContent = (NumMSAASamples > 1 ? false : bKeepDepthContent);

上述代码还揭示了平面反射在移动端的一种特殊渲染方式:Pixel Projected Reflection(PPR)。它的实现原理类似于SSR,但需要的数据更少,只需要场景颜色、深度缓冲和反射区域。它的核心步骤:

  • 计算场景颜色的所有像素在反射平面的镜像位置。
  • 测试像素的反射是否在反射区域内。
    • 光线投射到镜像像素位置。
    • 测试交点是否在反射区域内。
  • 如果找到相交点,计算像素在屏幕的镜像位置。
  • 在交点处写入镜像像素的颜色。

  • 如果反射区域内的交点被其它物体遮挡,则剔除此位置的反射。

PPR效果一览。

PPR可以在工程配置中设置:

UE在4.26在移动端渲染管线增加了延迟渲染分支,并在4.27做了改进和优化。移动端是否开启延迟着色的特性由以下代码决定:

// Engine\Source\Runtime\RenderCore\Private\RenderUtils.cpp

bool IsMobileDeferredShadingEnabled(const FStaticShaderPlatform Platform)
{
    // 禁用OpenGL的延迟着色.
    if (IsOpenGLPlatform(Platform))
    {
        // needs MRT framebuffer fetch or PLS
        return false;
    }

    // 控制台变量"r.Mobile.ShadingPath"要为1.
    static auto* MobileShadingPathCvar = IConsoleManager::Get().FindTConsoleVariableDataInt(TEXT("r.Mobile.ShadingPath"));
    return MobileShadingPathCvar->GetValueOnAnyThread() == 1;
}

简单地说就是非OpenGL图形API且控制台变量r.Mobile.ShadingPath设为1。

r.Mobile.ShadingPath不可在编辑器动态设置值,只能在项目工程根目录/Config/DefaultEngine.ini增加以下字段来开启:

[/Script/Engine.RendererSettings]

r.Mobile.ShadingPath=1

添加以上字段后,重启UE编辑器,等待shader编译完成即可预览移动端延迟着色效果。

以下是延迟渲染分支FMobileSceneRenderer::RenderDeferred的代码和解析:

FRHITexture* FMobileSceneRenderer::RenderDeferred(FRHICommandListImmediate& RHICmdList, const TArrayView<const FViewInfo*> ViewList, const FSortedLightSetSceneInfo& SortedLightSet)
{
    FSceneRenderTargets& SceneContext = FSceneRenderTargets::Get(RHICmdList);

    // 准备GBuffer.
    FRHITexture* ColorTargets[4] = {
        SceneContext.GetSceneColorSurface(),
        SceneContext.GetGBufferATexture().GetReference(),
        SceneContext.GetGBufferBTexture().GetReference(),
        SceneContext.GetGBufferCTexture().GetReference()
    };

    // RHI是否需要将GBuffer存储到GPU的系统内存中,并在单独的渲染通道中进行着色.
    ERenderTargetActions GBufferAction = bRequiresMultiPass ? ERenderTargetActions::Clear_Store : ERenderTargetActions::Clear_DontStore;
    EDepthStencilTargetActions DepthAction = bKeepDepthContent ? EDepthStencilTargetActions::ClearDepthStencil_StoreDepthStencil : EDepthStencilTargetActions::ClearDepthStencil_DontStoreDepthStencil;

    // RT的load/store动作.
    ERenderTargetActions ColorTargetsAction[4] = {ERenderTargetActions::Clear_Store, GBufferAction, GBufferAction, GBufferAction};
    if (bIsFullPrepassEnabled)
    {
        ERenderTargetActions DepthTarget = MakeRenderTargetActions(ERenderTargetLoadAction::ELoad, GetStoreAction(GetDepthActions(DepthAction)));
        ERenderTargetActions StencilTarget = MakeRenderTargetActions(ERenderTargetLoadAction::ELoad, GetStoreAction(GetStencilActions(DepthAction)));
        DepthAction = MakeDepthStencilTargetActions(DepthTarget, StencilTarget);
    }

    FRHIRenderPassInfo BasePassInfo = FRHIRenderPassInfo();
    int32 ColorTargetIndex = 0;
    for (; ColorTargetIndex < UE_ARRAY_COUNT(ColorTargets); ++ColorTargetIndex)
    {
        BasePassInfo.ColorRenderTargets[ColorTargetIndex].RenderTarget = ColorTargets[ColorTargetIndex];
        BasePassInfo.ColorRenderTargets[ColorTargetIndex].ResolveTarget = nullptr;
        BasePassInfo.ColorRenderTargets[ColorTargetIndex].ArraySlice = -1;
        BasePassInfo.ColorRenderTargets[ColorTargetIndex].MipIndex = 0;
        BasePassInfo.ColorRenderTargets[ColorTargetIndex].Action = ColorTargetsAction[ColorTargetIndex];
    }

    if (MobileRequiresSceneDepthAux(ShaderPlatform))
    {
        BasePassInfo.ColorRenderTargets[ColorTargetIndex].RenderTarget = SceneContext.SceneDepthAux->GetRenderTargetItem().ShaderResourceTexture.GetReference();
        BasePassInfo.ColorRenderTargets[ColorTargetIndex].ResolveTarget = nullptr;
        BasePassInfo.ColorRenderTargets[ColorTargetIndex].ArraySlice = -1;
        BasePassInfo.ColorRenderTargets[ColorTargetIndex].MipIndex = 0;
        BasePassInfo.ColorRenderTargets[ColorTargetIndex].Action = GBufferAction;
        ColorTargetIndex++;
    }

    BasePassInfo.DepthStencilRenderTarget.DepthStencilTarget = SceneContext.GetSceneDepthSurface();
    BasePassInfo.DepthStencilRenderTarget.ResolveTarget = nullptr;
    BasePassInfo.DepthStencilRenderTarget.Action = DepthAction;
    BasePassInfo.DepthStencilRenderTarget.ExclusiveDepthStencil = FExclusiveDepthStencil::DepthWrite_StencilWrite;

    BasePassInfo.SubpassHint = ESubpassHint::DeferredShadingSubpass;
    if (!bIsFullPrepassEnabled)
    {
        BasePassInfo.NumOcclusionQueries = ComputeNumOcclusionQueriesToBatch();
        BasePassInfo.bOcclusionQueries = BasePassInfo.NumOcclusionQueries != 0;
    }
    BasePassInfo.ShadingRateTexture = nullptr;
    BasePassInfo.bIsMSAA = false;
    BasePassInfo.MultiViewCount = 0;

    RHICmdList.BeginRenderPass(BasePassInfo, TEXT("BasePassRendering"));

    if (GIsEditor && !Views[0].bIsSceneCapture)
    {
        DrawClearQuad(RHICmdList, Views[0].BackgroundColor);
    }

    // 深度PrePass
    if (!bIsFullPrepassEnabled)
    {
        RHICmdList.SetCurrentStat(GET_STATID(STAT_CLM_MobilePrePass));
        // Depth pre-pass
        RenderPrePass(RHICmdList);
    }

    // BasePass: 不透明和镂空物体.
    RHICmdList.SetCurrentStat(GET_STATID(STAT_CLMM_Opaque));
    RenderMobileBasePass(RHICmdList, ViewList);
    RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);

    // 遮挡剔除.
    if (!bIsFullPrepassEnabled)
    {
        // Issue occlusion queries
        RHICmdList.SetCurrentStat(GET_STATID(STAT_CLMM_Occlusion));
        RenderOcclusion(RHICmdList);
        RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);
    }

    // 非多Pass模式
    if (!bRequiresMultiPass)
    {
        // 下个子Pass: SSceneColor + GBuffer写入, SceneDepth只读.
        RHICmdList.NextSubpass();

        // 渲染贴花.
        if (ViewFamily.EngineShowFlags.Decals)
        {
            CSV_SCOPED_TIMING_STAT_EXCLUSIVE(RenderDecals);
            RenderDecals(RHICmdList);
        }

        // 下个子Pass: SceneColor写入, SceneDepth只读
        RHICmdList.NextSubpass();

        // 延迟光照着色.
        MobileDeferredShadingPass(RHICmdList, *Scene, ViewList, SortedLightSet);

        // 绘制半透明.
        if (ViewFamily.EngineShowFlags.Translucency)
        {
            CSV_SCOPED_TIMING_STAT_EXCLUSIVE(RenderTranslucency);
            SCOPE_CYCLE_COUNTER(STAT_TranslucencyDrawTime);
            RenderTranslucency(RHICmdList, ViewList);
            FRHICommandListExecutor::GetImmediateCommandList().PollOcclusionQueries();
            RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);
        }

        // 结束渲染Pass.
        RHICmdList.EndRenderPass();
    }
    // 多Pass模式(PC设备模拟的移动端).
    else
    {
        // 结束子pass.
        RHICmdList.NextSubpass();
        RHICmdList.NextSubpass();
        RHICmdList.EndRenderPass();

        // SceneColor + GBuffer write, SceneDepth is read only
        {
            for (int32 Index = 0; Index < UE_ARRAY_COUNT(ColorTargets); ++Index)
            {
                BasePassInfo.ColorRenderTargets[Index].Action = ERenderTargetActions::Load_Store;
            }
            BasePassInfo.DepthStencilRenderTarget.Action = EDepthStencilTargetActions::LoadDepthStencil_StoreDepthStencil;
            BasePassInfo.DepthStencilRenderTarget.ExclusiveDepthStencil = FExclusiveDepthStencil::DepthRead_StencilRead;
            BasePassInfo.SubpassHint = ESubpassHint::None;
            BasePassInfo.NumOcclusionQueries = 0;
            BasePassInfo.bOcclusionQueries = false;

            RHICmdList.BeginRenderPass(BasePassInfo, TEXT("AfterBasePass"));

            // 渲染贴花.
            if (ViewFamily.EngineShowFlags.Decals)
            {
                CSV_SCOPED_TIMING_STAT_EXCLUSIVE(RenderDecals);
                RenderDecals(RHICmdList);
            }

            RHICmdList.EndRenderPass();
        }

        // SceneColor write, SceneDepth is read only
        {
            FRHIRenderPassInfo ShadingPassInfo(
                SceneContext.GetSceneColorSurface(),
                ERenderTargetActions::Load_Store,
                nullptr,
                SceneContext.GetSceneDepthSurface(),
                EDepthStencilTargetActions::LoadDepthStencil_StoreDepthStencil,
                nullptr,
                nullptr,
                VRSRB_Passthrough,
                FExclusiveDepthStencil::DepthRead_StencilWrite
            );
            ShadingPassInfo.NumOcclusionQueries = 0;
            ShadingPassInfo.bOcclusionQueries = false;

            RHICmdList.BeginRenderPass(ShadingPassInfo, TEXT("MobileShadingPass"));

            // 延迟光照着色.
            MobileDeferredShadingPass(RHICmdList, *Scene, ViewList, SortedLightSet);

            // 绘制半透明.
            if (ViewFamily.EngineShowFlags.Translucency)
            {
                CSV_SCOPED_TIMING_STAT_EXCLUSIVE(RenderTranslucency);
                SCOPE_CYCLE_COUNTER(STAT_TranslucencyDrawTime);
                RenderTranslucency(RHICmdList, ViewList);
                FRHICommandListExecutor::GetImmediateCommandList().PollOcclusionQueries();
                RHICmdList.ImmediateFlush(EImmediateFlushType::DispatchToRHIThread);
            }

            RHICmdList.EndRenderPass();
        }
    }

    return ColorTargets[0];
}

由上面可知,移动端的延迟渲染管线和PC比较类似,先渲染BasePass,获得GBuffer几何信息,再执行光照计算。它们的流程图如下:

stateDiagram-v2
[*] --> DrawClearQuad*

DrawClearQuad* --> RenderPrePass*
RenderPrePass* --> RenderMobileBasePass
RenderMobileBasePass --> RenderOcclusion*

RenderOcclusion* --> RenderDecals*
RenderDecals* --> MobileDeferredShadingPass*
MobileDeferredShadingPass* --> RenderTranslucency*
RenderTranslucency* --> PreTonemapMSAA*

PreTonemapMSAA* --> [*]

当然,也有和PC不一样的地方,最明显的是移动端使用了适配TB(D)R架构的SubPass渲染,使得移动端在渲染PrePass深度、BasePass、光照计算时,让场景颜色、深度、GBuffer等信息一直在On-Chip的缓冲区中,提升渲染效率,降低设备能耗。

12.3.3.1 MobileDeferredShadingPass

延迟渲染光照的过程由MobileDeferredShadingPass担当:

void MobileDeferredShadingPass(
    FRHICommandListImmediate& RHICmdList,
    const FScene& Scene,
    const TArrayView<const FViewInfo*> PassViews,
    const FSortedLightSetSceneInfo &SortedLightSet)
{
    SCOPED_DRAW_EVENT(RHICmdList, MobileDeferredShading);

    const FViewInfo& View0 = *PassViews[0];

    FSceneRenderTargets& SceneContext = FSceneRenderTargets::Get(RHICmdList);
    // 创建Uniform Buffer.
    FUniformBufferRHIRef PassUniformBuffer = CreateMobileSceneTextureUniformBuffer(RHICmdList);
    FUniformBufferStaticBindings GlobalUniformBuffers(PassUniformBuffer);
    SCOPED_UNIFORM_BUFFER_GLOBAL_BINDINGS(RHICmdList, GlobalUniformBuffers);
    // 设置视口.
    RHICmdList.SetViewport(View0.ViewRect.Min.X, View0.ViewRect.Min.Y, 0.0f, View0.ViewRect.Max.X, View0.ViewRect.Max.Y, 1.0f);

    // 光照的默认材质.
    FCachedLightMaterial DefaultMaterial;
    DefaultMaterial.MaterialProxy = UMaterial::GetDefaultMaterial(MD_LightFunction)->GetRenderProxy();
    DefaultMaterial.Material = DefaultMaterial.MaterialProxy->GetMaterialNoFallback(ERHIFeatureLevel::ES3_1);
    check(DefaultMaterial.Material);

    // 绘制平行光.
    RenderDirectLight(RHICmdList, Scene, View0, DefaultMaterial);

    if (GMobileUseClusteredDeferredShading == 0)
    {
        // 渲染非分簇的简单光源.
        RenderSimpleLights(RHICmdList, Scene, PassViews, SortedLightSet, DefaultMaterial);
    }

    // 渲染非分簇的局部光源.
    int32 NumLights = SortedLightSet.SortedLights.Num();
    int32 StandardDeferredStart = SortedLightSet.SimpleLightsEnd;
    if (GMobileUseClusteredDeferredShading != 0)
    {
        StandardDeferredStart = SortedLightSet.ClusteredSupportedEnd;
    }

    // 渲染局部光源.
    for (int32 LightIdx = StandardDeferredStart; LightIdx < NumLights; ++LightIdx)
    {
        const FSortedLightSceneInfo& SortedLight = SortedLightSet.SortedLights[LightIdx];
        const FLightSceneInfo& LightSceneInfo = *SortedLight.LightSceneInfo;
        RenderLocalLight(RHICmdList, Scene, View0, LightSceneInfo, DefaultMaterial);
    }
}

下面继续分析渲染不同类型光源的接口:

// Engine\Source\Runtime\Renderer\Private\MobileDeferredShadingPass.cpp

// 渲染平行光
static void RenderDirectLight(FRHICommandListImmediate& RHICmdList, const FScene& Scene, const FViewInfo& View, const FCachedLightMaterial& DefaultLightMaterial)
{
    FSceneRenderTargets& SceneContext = FSceneRenderTargets::Get(RHICmdList);

    // 查找第一个平行光.
    FLightSceneInfo* DirectionalLight = nullptr;
    for (int32 ChannelIdx = 0; ChannelIdx < UE_ARRAY_COUNT(Scene.MobileDirectionalLights) && !DirectionalLight; ChannelIdx++)
    {
        DirectionalLight = Scene.MobileDirectionalLights[ChannelIdx];
    }

    // 渲染状态.
    FGraphicsPipelineStateInitializer GraphicsPSOInit;
    RHICmdList.ApplyCachedRenderTargets(GraphicsPSOInit);
    // 增加自发光到SceneColor.
    GraphicsPSOInit.BlendState = TStaticBlendState<CW_RGB, BO_Add, BF_One, BF_One>::GetRHI();
    GraphicsPSOInit.RasterizerState = TStaticRasterizerState<>::GetRHI();
    // 只绘制默认光照模型(MSM_DefaultLit)的像素.
    uint8 StencilRef = GET_STENCIL_MOBILE_SM_MASK(MSM_DefaultLit);
    GraphicsPSOInit.DepthStencilState = TStaticDepthStencilState<
                                        false, CF_Always,
                                        true, CF_Equal, SO_Keep, SO_Keep, SO_Keep,
                                        false, CF_Always, SO_Keep, SO_Keep, SO_Keep,
                                        GET_STENCIL_MOBILE_SM_MASK(0x7), 0x00>::GetRHI(); // 4 bits for shading models

    // 处理VS.
    TShaderMapRef<FPostProcessVS> VertexShader(View.ShaderMap);

    const FMaterialRenderProxy* LightFunctionMaterialProxy = nullptr;
    if (View.Family->EngineShowFlags.LightFunctions && DirectionalLight)
    {
        LightFunctionMaterialProxy = DirectionalLight->Proxy->GetLightFunctionMaterial();
    }
    FMobileDirectLightFunctionPS::FPermutationDomain PermutationVector = FMobileDirectLightFunctionPS::BuildPermutationVector(View, DirectionalLight != nullptr);
    FCachedLightMaterial LightMaterial;
    TShaderRef<FMobileDirectLightFunctionPS> PixelShader;
    GetLightMaterial(DefaultLightMaterial, LightFunctionMaterialProxy, PermutationVector.ToDimensionValueId(), LightMaterial, PixelShader);

    GraphicsPSOInit.BoundShaderState.VertexDeclarationRHI = GFilterVertexDeclaration.VertexDeclarationRHI;
    GraphicsPSOInit.BoundShaderState.VertexShaderRHI = VertexShader.GetVertexShader();
    GraphicsPSOInit.BoundShaderState.PixelShaderRHI = PixelShader.GetPixelShader();
    GraphicsPSOInit.PrimitiveType = PT_TriangleList;
    SetGraphicsPipelineState(RHICmdList, GraphicsPSOInit);

    // 处理PS.
    FMobileDirectLightFunctionPS::FParameters PassParameters;
    PassParameters.Forward = View.ForwardLightingResources->ForwardLightDataUniformBuffer;
    PassParameters.MobileDirectionalLight = Scene.UniformBuffers.MobileDirectionalLightUniformBuffers[1];
    PassParameters.ReflectionCaptureData = Scene.UniformBuffers.ReflectionCaptureUniformBuffer;
    FReflectionUniformParameters ReflectionUniformParameters;
    SetupReflectionUniformParameters(View, ReflectionUniformParameters);
    PassParameters.ReflectionsParameters = CreateUniformBufferImmediate(ReflectionUniformParameters, UniformBuffer_SingleDraw);
    PassParameters.LightFunctionParameters = FVector4(1.0f, 1.0f, 0.0f, 0.0f);
    if (DirectionalLight)
    {
        const bool bUseMovableLight = DirectionalLight && !DirectionalLight->Proxy->HasStaticShadowing();
        PassParameters.LightFunctionParameters2 = FVector(DirectionalLight->Proxy->GetLightFunctionFadeDistance(), DirectionalLight->Proxy->GetLightFunctionDisabledBrightness(), bUseMovableLight ? 1.0f : 0.0f);
        const FVector Scale = DirectionalLight->Proxy->GetLightFunctionScale();
        // Switch x and z so that z of the user specified scale affects the distance along the light direction
        const FVector InverseScale = FVector(1.f / Scale.Z, 1.f / Scale.Y, 1.f / Scale.X);
        PassParameters.WorldToLight = DirectionalLight->Proxy->GetWorldToLight() * FScaleMatrix(FVector(InverseScale));
    }
    FMobileDirectLightFunctionPS::SetParameters(RHICmdList, PixelShader, View, LightMaterial.MaterialProxy, *LightMaterial.Material, PassParameters);

    RHICmdList.SetStencilRef(StencilRef);

    const FIntPoint TargetSize = SceneContext.GetBufferSizeXY();

    // 用全屏幕的矩形绘制.
    DrawRectangle(
        RHICmdList,
        0, 0,
        View.ViewRect.Width(), View.ViewRect.Height(),
        View.ViewRect.Min.X, View.ViewRect.Min.Y,
        View.ViewRect.Width(), View.ViewRect.Height(),
        FIntPoint(View.ViewRect.Width(), View.ViewRect.Height()),
        TargetSize,
        VertexShader);
}

// 渲染非分簇模式的简单光源.
static void RenderSimpleLights(
    FRHICommandListImmediate& RHICmdList,
    const FScene& Scene,
    const TArrayView<const FViewInfo*> PassViews,
    const FSortedLightSetSceneInfo &SortedLightSet,
    const FCachedLightMaterial& DefaultMaterial)
{
    const FSimpleLightArray& SimpleLights = SortedLightSet.SimpleLights;
    const int32 NumViews = PassViews.Num();
    const FViewInfo& View0 = *PassViews[0];

    // 处理VS.
    TShaderMapRef<TDeferredLightVS<true>> VertexShader(View0.ShaderMap);
    TShaderRef<FMobileRadialLightFunctionPS> PixelShaders[2];
    {
        const FMaterialShaderMap* MaterialShaderMap = DefaultMaterial.Material->GetRenderingThreadShaderMap();
        FMobileRadialLightFunctionPS::FPermutationDomain PermutationVector;
        PermutationVector.Set<FMobileRadialLightFunctionPS::FSpotLightDim>(false);
        PermutationVector.Set<FMobileRadialLightFunctionPS::FIESProfileDim>(false);
        PermutationVector.Set<FMobileRadialLightFunctionPS::FInverseSquaredDim>(false);
        PixelShaders[0] = MaterialShaderMap->GetShader<FMobileRadialLightFunctionPS>(PermutationVector);
        PermutationVector.Set<FMobileRadialLightFunctionPS::FInverseSquaredDim>(true);
        PixelShaders[1] = MaterialShaderMap->GetShader<FMobileRadialLightFunctionPS>(PermutationVector);
    }

    // 设置PSO.
    FGraphicsPipelineStateInitializer GraphicsPSOLight[2];
    {
        SetupSimpleLightPSO(RHICmdList, View0, VertexShader, PixelShaders[0], GraphicsPSOLight[0]);
        SetupSimpleLightPSO(RHICmdList, View0, VertexShader, PixelShaders[1], GraphicsPSOLight[1]);
    }

    // 设置模板缓冲.
    FGraphicsPipelineStateInitializer GraphicsPSOLightMask;
    {
        RHICmdList.ApplyCachedRenderTargets(GraphicsPSOLightMask);
        GraphicsPSOLightMask.PrimitiveType = PT_TriangleList;
        GraphicsPSOLightMask.BlendState = TStaticBlendStateWriteMask<CW_NONE, CW_NONE, CW_NONE, CW_NONE, CW_NONE, CW_NONE, CW_NONE, CW_NONE>::GetRHI();
        GraphicsPSOLightMask.RasterizerState = View0.bReverseCulling ? TStaticRasterizerState<FM_Solid, CM_CCW>::GetRHI() : TStaticRasterizerState<FM_Solid, CM_CW>::GetRHI();
        // set stencil to 1 where depth test fails
        GraphicsPSOLightMask.DepthStencilState = TStaticDepthStencilState<
            false, CF_DepthNearOrEqual,
            true, CF_Always, SO_Keep, SO_Replace, SO_Keep,
            false, CF_Always, SO_Keep, SO_Keep, SO_Keep,
            0x00, STENCIL_SANDBOX_MASK>::GetRHI();
        GraphicsPSOLightMask.BoundShaderState.VertexDeclarationRHI = GetVertexDeclarationFVector4();
        GraphicsPSOLightMask.BoundShaderState.VertexShaderRHI = VertexShader.GetVertexShader();
        GraphicsPSOLightMask.BoundShaderState.PixelShaderRHI = nullptr;
    }

    // 遍历所有简单光源列表, 执行着色计算.
    for (int32 LightIndex = 0; LightIndex < SimpleLights.InstanceData.Num(); LightIndex++)
    {
        const FSimpleLightEntry& SimpleLight = SimpleLights.InstanceData[LightIndex];
        for (int32 ViewIndex = 0; ViewIndex < NumViews; ViewIndex++)
        {
            const FViewInfo& View = *PassViews[ViewIndex];
            const FSimpleLightPerViewEntry& SimpleLightPerViewData = SimpleLights.GetViewDependentData(LightIndex, ViewIndex, NumViews);
            const FSphere LightBounds(SimpleLightPerViewData.Position, SimpleLight.Radius);

            if (NumViews > 1)
            {
                // set viewports only we we have more than one
                // otherwise it is set at the start of the pass
                RHICmdList.SetViewport(View.ViewRect.Min.X, View.ViewRect.Min.Y, 0.0f, View.ViewRect.Max.X, View.ViewRect.Max.Y, 1.0f);
            }

            // 渲染光源遮罩.
            SetGraphicsPipelineState(RHICmdList, GraphicsPSOLightMask);
            VertexShader->SetSimpleLightParameters(RHICmdList, View, LightBounds);
            RHICmdList.SetStencilRef(1);
            StencilingGeometry::DrawSphere(RHICmdList);

            // 渲染光源.
            FMobileRadialLightFunctionPS::FParameters PassParameters;
            FDeferredLightUniformStruct DeferredLightUniformsValue;
            SetupSimpleDeferredLightParameters(SimpleLight, SimpleLightPerViewData, DeferredLightUniformsValue);
            PassParameters.DeferredLightUniforms = TUniformBufferRef<FDeferredLightUniformStruct>::CreateUniformBufferImmediate(DeferredLightUniformsValue, EUniformBufferUsage::UniformBuffer_SingleFrame);
            PassParameters.IESTexture = GWhiteTexture->TextureRHI;
            PassParameters.IESTextureSampler = GWhiteTexture->SamplerStateRHI;
            if (SimpleLight.Exponent == 0)
            {
                SetGraphicsPipelineState(RHICmdList, GraphicsPSOLight[1]);
                FMobileRadialLightFunctionPS::SetParameters(RHICmdList, PixelShaders[1], View, DefaultMaterial.MaterialProxy, *DefaultMaterial.Material, PassParameters);
            }
            else
            {
                SetGraphicsPipelineState(RHICmdList, GraphicsPSOLight[0]);
                FMobileRadialLightFunctionPS::SetParameters(RHICmdList, PixelShaders[0], View, DefaultMaterial.MaterialProxy, *DefaultMaterial.Material, PassParameters);
            }
            VertexShader->SetSimpleLightParameters(RHICmdList, View, LightBounds);

            // 只绘制默认光照模型(MSM_DefaultLit)的像素.
            uint8 StencilRef = GET_STENCIL_MOBILE_SM_MASK(MSM_DefaultLit);
            RHICmdList.SetStencilRef(StencilRef);

            // 用球体渲染光源(点光源和聚光灯), 以快速剔除光源影响之外的像素.
            StencilingGeometry::DrawSphere(RHICmdList);
        }
    }
}

// 渲染局部光源.
static void RenderLocalLight(
    FRHICommandListImmediate& RHICmdList,
    const FScene& Scene,
    const FViewInfo& View,
    const FLightSceneInfo& LightSceneInfo,
    const FCachedLightMaterial& DefaultLightMaterial)
{
    if (!LightSceneInfo.ShouldRenderLight(View))
    {
        return;
    }

    // 忽略非局部光源(光源和聚光灯之外的光源).
    const uint8 LightType = LightSceneInfo.Proxy->GetLightType();
    const bool bIsSpotLight = LightType == LightType_Spot;
    const bool bIsPointLight = LightType == LightType_Point;
    if (!bIsSpotLight && !bIsPointLight)
    {
        return;
    }

    // 绘制光源模板.
    if (GMobileUseLightStencilCulling != 0)
    {
        RenderLocalLight_StencilMask(RHICmdList, Scene, View, LightSceneInfo);
    }

    // 处理IES光照.
    bool bUseIESTexture = false;
    FTexture* IESTextureResource = GWhiteTexture;
    if (View.Family->EngineShowFlags.TexturedLightProfiles && LightSceneInfo.Proxy->GetIESTextureResource())
    {
        IESTextureResource = LightSceneInfo.Proxy->GetIESTextureResource();
        bUseIESTexture = true;
    }

    FGraphicsPipelineStateInitializer GraphicsPSOInit;
    RHICmdList.ApplyCachedRenderTargets(GraphicsPSOInit);
    GraphicsPSOInit.BlendState = TStaticBlendState<CW_RGBA, BO_Add, BF_One, BF_One, BO_Add, BF_One, BF_One>::GetRHI();
    GraphicsPSOInit.PrimitiveType = PT_TriangleList;
    const FSphere LightBounds = LightSceneInfo.Proxy->GetBoundingSphere();

    // 设置光源光栅化和深度状态.
    if (GMobileUseLightStencilCulling != 0)
    {
        SetLocalLightRasterizerAndDepthState_StencilMask(GraphicsPSOInit, View);
    }
    else
    {
        SetLocalLightRasterizerAndDepthState(GraphicsPSOInit, View, LightBounds);
    }

    // 设置VS
    TShaderMapRef<TDeferredLightVS<true>> VertexShader(View.ShaderMap);

    const FMaterialRenderProxy* LightFunctionMaterialProxy = nullptr;
    if (View.Family->EngineShowFlags.LightFunctions)
    {
        LightFunctionMaterialProxy = LightSceneInfo.Proxy->GetLightFunctionMaterial();
    }
    FMobileRadialLightFunctionPS::FPermutationDomain PermutationVector;
    PermutationVector.Set<FMobileRadialLightFunctionPS::FSpotLightDim>(bIsSpotLight);
    PermutationVector.Set<FMobileRadialLightFunctionPS::FInverseSquaredDim>(LightSceneInfo.Proxy->IsInverseSquared());
    PermutationVector.Set<FMobileRadialLightFunctionPS::FIESProfileDim>(bUseIESTexture);
    FCachedLightMaterial LightMaterial;
    TShaderRef<FMobileRadialLightFunctionPS> PixelShader;
    GetLightMaterial(DefaultLightMaterial, LightFunctionMaterialProxy, PermutationVector.ToDimensionValueId(), LightMaterial, PixelShader);

    GraphicsPSOInit.BoundShaderState.VertexDeclarationRHI = GetVertexDeclarationFVector4();
    GraphicsPSOInit.BoundShaderState.VertexShaderRHI = VertexShader.GetVertexShader();
    GraphicsPSOInit.BoundShaderState.PixelShaderRHI = PixelShader.GetPixelShader();
    SetGraphicsPipelineState(RHICmdList, GraphicsPSOInit);

    VertexShader->SetParameters(RHICmdList, View, &LightSceneInfo);

    // 设置PS.
    FMobileRadialLightFunctionPS::FParameters PassParameters;
    PassParameters.DeferredLightUniforms = TUniformBufferRef<FDeferredLightUniformStruct>::CreateUniformBufferImmediate(GetDeferredLightParameters(View, LightSceneInfo), EUniformBufferUsage::UniformBuffer_SingleFrame);
    PassParameters.IESTexture = IESTextureResource->TextureRHI;
    PassParameters.IESTextureSampler = IESTextureResource->SamplerStateRHI;
    const float TanOuterAngle = bIsSpotLight ? FMath::Tan(LightSceneInfo.Proxy->GetOuterConeAngle()) : 1.0f;
    PassParameters.LightFunctionParameters = FVector4(TanOuterAngle, 1.0f /*ShadowFadeFraction*/, bIsSpotLight ? 1.0f : 0.0f, bIsPointLight ? 1.0f : 0.0f);
    PassParameters.LightFunctionParameters2 = FVector(LightSceneInfo.Proxy->GetLightFunctionFadeDistance(), LightSceneInfo.Proxy->GetLightFunctionDisabledBrightness(),    0.0f);
    const FVector Scale = LightSceneInfo.Proxy->GetLightFunctionScale();
    // Switch x and z so that z of the user specified scale affects the distance along the light direction
    const FVector InverseScale = FVector(1.f / Scale.Z, 1.f / Scale.Y, 1.f / Scale.X);
    PassParameters.WorldToLight = LightSceneInfo.Proxy->GetWorldToLight() * FScaleMatrix(FVector(InverseScale));
    FMobileRadialLightFunctionPS::SetParameters(RHICmdList, PixelShader, View, LightMaterial.MaterialProxy, *LightMaterial.Material, PassParameters);

    // 只绘制默认光照模型(MSM_DefaultLit)的像素.
    uint8 StencilRef = GET_STENCIL_MOBILE_SM_MASK(MSM_DefaultLit);
    RHICmdList.SetStencilRef(StencilRef);

    // 点光源用球体绘制.
    if (LightType == LightType_Point)
    {
        StencilingGeometry::DrawSphere(RHICmdList);
    }
    // 聚光灯用锥体绘制.
    else // LightType_Spot
    {
        StencilingGeometry::DrawCone(RHICmdList);
    }
}

绘制光源时,按光源类型划分为三个步骤:平行光、非分簇简单光源、局部光源(点光源和聚光灯)。需要注意的是,移动端只支持默认光照模型(MSM_DefaultLit)的计算,其它高级光照模型(头发、次表面散射、清漆、眼睛、布料等)暂不支持。

绘制平行光时,最多只能绘制1个,采用的是全屏幕矩形绘制,支持若干级CSM阴影。

绘制非分簇简单光源时,无论是点光源还是聚光灯,都采用球体绘制,不支持阴影。

绘制局部光源时,会复杂许多,先绘制局部光源模板缓冲,再设置光栅化和深度状态,最后才绘制光源。其中点光源采用球体绘制,不支持阴影;聚光灯采用锥体绘制,可以支持阴影,默认情况下,聚光灯不支持动态光影计算,需要在工程配置中开启:

此外,是否开启模板剔除光源不相交的像素由GMobileUseLightStencilCulling决定,而GMobileUseLightStencilCulling又由r.Mobile.UseLightStencilCulling决定,默认为1(即开启状态)。渲染光源的模板缓冲代码如下:

static void RenderLocalLight_StencilMask(FRHICommandListImmediate& RHICmdList, const FScene& Scene, const FViewInfo& View, const FLightSceneInfo& LightSceneInfo)
{
    const uint8 LightType = LightSceneInfo.Proxy->GetLightType();

    FGraphicsPipelineStateInitializer GraphicsPSOInit;
    // 应用缓存好的RT(颜色/深度等).
    RHICmdList.ApplyCachedRenderTargets(GraphicsPSOInit);
    GraphicsPSOInit.PrimitiveType = PT_TriangleList;
    // 禁用所有RT的写操作.
    GraphicsPSOInit.BlendState = TStaticBlendStateWriteMask<CW_NONE, CW_NONE, CW_NONE, CW_NONE, CW_NONE, CW_NONE, CW_NONE, CW_NONE>::GetRHI();
    GraphicsPSOInit.RasterizerState = View.bReverseCulling ? TStaticRasterizerState<FM_Solid, CM_CCW>::GetRHI() : TStaticRasterizerState<FM_Solid, CM_CW>::GetRHI();
    // 如果深度测试失败, 则写入模板缓冲值为1.
    GraphicsPSOInit.DepthStencilState = TStaticDepthStencilState<
        false, CF_DepthNearOrEqual,
        true, CF_Always, SO_Keep, SO_Replace, SO_Keep,
        false, CF_Always, SO_Keep, SO_Keep, SO_Keep,
        0x00,
        // 注意只写入Pass专用的沙盒(SANBOX)位, 即模板缓冲的索引为0的位.
        STENCIL_SANDBOX_MASK>::GetRHI();

    // 绘制光源模板的VS是TDeferredLightVS.
    TShaderMapRef<TDeferredLightVS<true> > VertexShader(View.ShaderMap);
    GraphicsPSOInit.BoundShaderState.VertexDeclarationRHI = GetVertexDeclarationFVector4();
    GraphicsPSOInit.BoundShaderState.VertexShaderRHI = VertexShader.GetVertexShader();
    // PS为空.
    GraphicsPSOInit.BoundShaderState.PixelShaderRHI = nullptr;

    SetGraphicsPipelineState(RHICmdList, GraphicsPSOInit);
    VertexShader->SetParameters(RHICmdList, View, &LightSceneInfo);
    // 模板值为1.
    RHICmdList.SetStencilRef(1);

    // 根据不同光源用不同形状绘制.
    if (LightType == LightType_Point)
    {
        StencilingGeometry::DrawSphere(RHICmdList);
    }
    else // LightType_Spot
    {
        StencilingGeometry::DrawCone(RHICmdList);
    }
}

每个局部光源首先绘制光源范围内的Mask,再计算通过了Stencil测试(Early-Z)的像素的光照。具体的剖析过程以下图的聚光灯为例:

上:场景中一盏等待渲染的聚光灯;中:利用模板Pass绘制出的模板Mask(白色区域),标记了屏幕空间中和聚光灯形状重叠且深度更近的像素 ;下:对有效像素进行光照计算后的效果。

对有效像素进行光照计算时,使用的DepthStencil状态如下:

翻译成文字就是,执行光照的像素必须在光源形状体之内,光源形状之外的像素会被剔除。模板Pass标记的是比光源形状深度更近的像素(光源形状体之外的像素),光源绘制Pass通过模板测试剔除模板Pass标记的像素,然后再通过深度测试找出在光源形状体内的像素,从而提升光照计算效率。

移动端的这种光源模板裁剪(Light Stencil Culling)技术和Siggraph2020的Unity演讲Deferred Shading in Unity URP提及的基于模板的光照计算相似(思想一致,但做法可能不完全一样)。该论文还提出了更加契合光源形状的几何体模拟:

以及对比了各种光源计算方法在PC和移动端的性能,下面是Mali GPU的对比图:

Mali Gpu在使用不同光照渲染技术的性能对比,可见在移动端,基于模板裁剪的光照算法要优于常规和分块算法。

值得一提的是,光源模板裁剪技术结合GPU的Early-Z技术,将极大提升光照渲染性能。而当前主流的移动端GPU都支持Early-Z技术,也为光源模板裁剪的应用奠定了基础。

UE目前实现的光源裁剪算法兴许还有改进的空间,比如背向光源的像素(下图红框所示)其实也是可以不计算的。(但如何快速有效地找到背向光源的像素又是一个问题)

12.3.3.2 MobileBasePassShader

本节主要阐述移动端BasePass涉及的shader,包括VS和PS。先看VS:

// Engine\Shaders\Private\MobileBasePassVertexShader.usf

(......)

struct FMobileShadingBasePassVSToPS
{
    FVertexFactoryInterpolantsVSToPS FactoryInterpolants;
    FMobileBasePassInterpolantsVSToPS BasePassInterpolants;
    float4 Position : SV_POSITION;
};

#define FMobileShadingBasePassVSOutput FMobileShadingBasePassVSToPS
#define VertexFactoryGetInterpolants VertexFactoryGetInterpolantsVSToPS

// VS主入口.
void Main(
    FVertexFactoryInput Input
    , out FMobileShadingBasePassVSOutput Output
#if INSTANCED_STEREO
    , uint InstanceId : SV_InstanceID
    , out uint LayerIndex : SV_RenderTargetArrayIndex
#elif MOBILE_MULTI_VIEW
    , in uint ViewId : SV_ViewID
#endif
    )
{
// 立体视图模式.
#if INSTANCED_STEREO
    const uint EyeIndex = GetEyeIndex(InstanceId);
    ResolvedView = ResolveView(EyeIndex);
    LayerIndex = EyeIndex;
    Output.BasePassInterpolants.MultiViewId = float(EyeIndex);
// 多视图模式.
#elif MOBILE_MULTI_VIEW
    #if COMPILER_GLSL_ES3_1
        const int MultiViewId = int(ViewId);
        ResolvedView = ResolveView(uint(MultiViewId));
        Output.BasePassInterpolants.MultiViewId = float(MultiViewId);
    #else
        ResolvedView = ResolveView(ViewId);
        Output.BasePassInterpolants.MultiViewId = float(ViewId);
    #endif
#else
    ResolvedView = ResolveView();
#endif

    // 初始化打包的插值数据.
#if PACK_INTERPOLANTS
    float4 PackedInterps[NUM_VF_PACKED_INTERPOLANTS];
    UNROLL
    for(int i = 0; i < NUM_VF_PACKED_INTERPOLANTS; ++i)
    {
        PackedInterps[i] = 0;
    }
#endif

    // 处理顶点工厂数据.
    FVertexFactoryIntermediates VFIntermediates = GetVertexFactoryIntermediates(Input);
    float4 WorldPositionExcludingWPO = VertexFactoryGetWorldPosition(Input, VFIntermediates);
    float4 WorldPosition = WorldPositionExcludingWPO;

    // 获取材质的顶点数据, 处理坐标等.
    half3x3 TangentToLocal = VertexFactoryGetTangentToLocal(Input, VFIntermediates);
    FMaterialVertexParameters VertexParameters = GetMaterialVertexParameters(Input, VFIntermediates, WorldPosition.xyz, TangentToLocal);

    half3 WorldPositionOffset = GetMaterialWorldPositionOffset(VertexParameters);

    WorldPosition.xyz += WorldPositionOffset;

    float4 RasterizedWorldPosition = VertexFactoryGetRasterizedWorldPosition(Input, VFIntermediates, WorldPosition);
    Output.Position = mul(RasterizedWorldPosition, ResolvedView.TranslatedWorldToClip);
    Output.BasePassInterpolants.PixelPosition = WorldPosition;

#if USE_WORLD_POSITION_EXCLUDING_SHADER_OFFSETS
    Output.BasePassInterpolants.PixelPositionExcludingWPO = WorldPositionExcludingWPO.xyz;
#endif

    // 裁剪面.
#if USE_PS_CLIP_PLANE
    Output.BasePassInterpolants.OutClipDistance = dot(ResolvedView.GlobalClippingPlane, float4(WorldPosition.xyz - ResolvedView.PreViewTranslation.xyz, 1));
#endif

    // 顶点雾.
#if USE_VERTEX_FOG
    float4 VertexFog = CalculateHeightFog(WorldPosition.xyz - ResolvedView.TranslatedWorldCameraOrigin);

    #if PROJECT_SUPPORT_SKY_ATMOSPHERE && MATERIAL_IS_SKY==0 // Do not apply aerial perpsective on sky materials
        if (ResolvedView.SkyAtmosphereApplyCameraAerialPerspectiveVolume > 0.0f)
        {
            const float OneOverPreExposure = USE_PREEXPOSURE ? ResolvedView.OneOverPreExposure : 1.0f;
            // Sample the aerial perspective (AP). It is also blended under the VertexFog parameter.
            VertexFog = GetAerialPerspectiveLuminanceTransmittanceWithFogOver(
                ResolvedView.RealTimeReflectionCapture, ResolvedView.SkyAtmosphereCameraAerialPerspectiveVolumeSizeAndInvSize,
                Output.Position, WorldPosition.xyz*CM_TO_SKY_UNIT, ResolvedView.TranslatedWorldCameraOrigin*CM_TO_SKY_UNIT,
                View.CameraAerialPerspectiveVolume, View.CameraAerialPerspectiveVolumeSampler,
                ResolvedView.SkyAtmosphereCameraAerialPerspectiveVolumeDepthResolutionInv,
                ResolvedView.SkyAtmosphereCameraAerialPerspectiveVolumeDepthResolution,
                ResolvedView.SkyAtmosphereAerialPerspectiveStartDepthKm,
                ResolvedView.SkyAtmosphereCameraAerialPerspectiveVolumeDepthSliceLengthKm,
                ResolvedView.SkyAtmosphereCameraAerialPerspectiveVolumeDepthSliceLengthKmInv,
                OneOverPreExposure, VertexFog);
        }
    #endif

    #if PACK_INTERPOLANTS
        PackedInterps[0] = VertexFog;
    #else
        Output.BasePassInterpolants.VertexFog = VertexFog;
    #endif // PACK_INTERPOLANTS
#endif // USE_VERTEX_FOG

    (......)

    // 获取待插值的数据.
    Output.FactoryInterpolants = VertexFactoryGetInterpolants(Input, VFIntermediates, VertexParameters);

    Output.BasePassInterpolants.PixelPosition.w = Output.Position.w;

    // 打包插值数据.
#if PACK_INTERPOLANTS
    VertexFactoryPackInterpolants(Output.FactoryInterpolants, PackedInterps);
#endif // PACK_INTERPOLANTS

#if !OUTPUT_MOBILE_HDR && COMPILER_GLSL_ES3_1
    Output.Position.y *= -1;
#endif
}

以上可知,视图实例会根据立体绘制、多视图和普通模式不同而不同处理。支持顶点雾,但默认是关闭的,需要在工程配置内开启。

存在打包插值模式,为了压缩VS到PS之间的插值消耗和带宽。是否开启由宏PACK_INTERPOLANTS决定,它的定义如下:

// Engine\Shaders\Private\MobileBasePassCommon.ush

#define PACK_INTERPOLANTS (USE_VERTEX_FOG && NUM_VF_PACKED_INTERPOLANTS > 0 && (ES3_1_PROFILE))

也就是说,只有开启顶点雾、存在顶点工厂打包插值数据且是OpenGLES3.1着色平台才开启打包插值的特性。相比PC端的BasePass的VS,移动端的做了大量的简化,可以简单地认为只是PC端的一个很小的子集。下面继续分析PS:

// Engine\Shaders\Private\MobileBasePassVertexShader.usf

#include "Common.ush"

// 各类宏定义.
#define MobileSceneTextures MobileBasePass.SceneTextures
#define EyeAdaptationStruct MobileBasePass

(......)

// 最接近被渲染对象的场景的预归一化捕获(完全粗糙材质不支持)
#if !FULLY_ROUGH
    #if HQ_REFLECTIONS
    #define MAX_HQ_REFLECTIONS 3
    TextureCube ReflectionCubemap0;
    SamplerState ReflectionCubemapSampler0;
    TextureCube ReflectionCubemap1;
    SamplerState ReflectionCubemapSampler1;
    TextureCube ReflectionCubemap2;
    SamplerState ReflectionCubemapSampler2;
    // x,y,z - inverted average brightness for 0, 1, 2; w - sky cube texture max mips.
    float4 ReflectionAverageBrigtness;
    float4 ReflectanceMaxValueRGBM;
    float4 ReflectionPositionsAndRadii[MAX_HQ_REFLECTIONS];
        #if ALLOW_CUBE_REFLECTIONS
        float4x4 CaptureBoxTransformArray[MAX_HQ_REFLECTIONS];
        float4 CaptureBoxScalesArray[MAX_HQ_REFLECTIONS];
        #endif
    #endif
#endif

// 反射球/IBL等接口.
half4 GetPlanarReflection(float3 WorldPosition, half3 WorldNormal, half Roughness);
half MobileComputeMixingWeight(half IndirectIrradiance, half AverageBrightness, half Roughness);
half3 GetLookupVectorForBoxCaptureMobile(half3 ReflectionVector, ...);
half3 GetLookupVectorForSphereCaptureMobile(half3 ReflectionVector, ...);
void GatherSpecularIBL(FMaterialPixelParameters MaterialParameters, ...);
void BlendReflectionCaptures(FMaterialPixelParameters MaterialParameters, ...)
half3 GetImageBasedReflectionLighting(FMaterialPixelParameters MaterialParameters, ...);

// 其它接口.
half3 FrameBufferBlendOp(half4 Source);
bool UseCSM();
void ApplyPixelDepthOffsetForMobileBasePass(inout FMaterialPixelParameters MaterialParameters, FPixelMaterialInputs PixelMaterialInputs, out float OutDepth);

// 累积动态点光源.
#if MAX_DYNAMIC_POINT_LIGHTS > 0
void AccumulateLightingOfDynamicPointLight(
    FMaterialPixelParameters MaterialParameters,
    FMobileShadingModelContext ShadingModelContext,
    FGBufferData GBuffer,
    float4 LightPositionAndInvRadius,
    float4 LightColorAndFalloffExponent,
    float4 SpotLightDirectionAndSpecularScale,
    float4 SpotLightAnglesAndSoftTransitionScaleAndLightShadowType,
    #if SUPPORT_SPOTLIGHTS_SHADOW
    FPCFSamplerSettings Settings,
    float4 SpotLightShadowSharpenAndShadowFadeFraction,
    float4 SpotLightShadowmapMinMax,
    float4x4 SpotLightShadowWorldToShadowMatrix,
    #endif
    inout half3 Color)
{
    uint LightShadowType = SpotLightAnglesAndSoftTransitionScaleAndLightShadowType.w;
    float FadedShadow = 1.0f;

    // 计算聚光灯阴影.
#if SUPPORT_SPOTLIGHTS_SHADOW
    if ((LightShadowType & LightShadowType_Shadow) == LightShadowType_Shadow)
    {

        float4 HomogeneousShadowPosition = mul(float4(MaterialParameters.AbsoluteWorldPosition, 1), SpotLightShadowWorldToShadowMatrix);
        float2 ShadowUVs = HomogeneousShadowPosition.xy / HomogeneousShadowPosition.w;
        if (all(ShadowUVs >= SpotLightShadowmapMinMax.xy && ShadowUVs <= SpotLightShadowmapMinMax.zw))
        {
            // Clamp pixel depth in light space for shadowing opaque, because areas of the shadow depth buffer that weren't rendered to will have been cleared to 1
            // We want to force the shadow comparison to result in 'unshadowed' in that case, regardless of whether the pixel being shaded is in front or behind that plane
            float LightSpacePixelDepthForOpaque = min(HomogeneousShadowPosition.z, 0.99999f);
            Settings.SceneDepth = LightSpacePixelDepthForOpaque;
            Settings.TransitionScale = SpotLightAnglesAndSoftTransitionScaleAndLightShadowType.z;

            half Shadow = MobileShadowPCF(ShadowUVs, Settings);

            Shadow = saturate((Shadow - 0.5) * SpotLightShadowSharpenAndShadowFadeFraction.x + 0.5);

            FadedShadow = lerp(1.0f, Square(Shadow), SpotLightShadowSharpenAndShadowFadeFraction.y);
        }
    }
#endif

    // 计算光照.
    if ((LightShadowType & ValidLightType) != 0)
    {
        float3 ToLight = LightPositionAndInvRadius.xyz - MaterialParameters.AbsoluteWorldPosition;
        float DistanceSqr = dot(ToLight, ToLight);
        float3 L = ToLight * rsqrt(DistanceSqr);
        half3 PointH = normalize(MaterialParameters.CameraVector + L);

        half PointNoL = max(0, dot(MaterialParameters.WorldNormal, L));
        half PointNoH = max(0, dot(MaterialParameters.WorldNormal, PointH));

        // 计算光源的衰减.
        float Attenuation;
        if (LightColorAndFalloffExponent.w == 0)
        {
            // Sphere falloff (technically just 1/d2 but this avoids inf)
            Attenuation = 1 / (DistanceSqr + 1);

            float LightRadiusMask = Square(saturate(1 - Square(DistanceSqr * (LightPositionAndInvRadius.w * LightPositionAndInvRadius.w))));
            Attenuation *= LightRadiusMask;
        }
        else
        {
            Attenuation = RadialAttenuation(ToLight * LightPositionAndInvRadius.w, LightColorAndFalloffExponent.w);
        }

#if PROJECT_MOBILE_ENABLE_MOVABLE_SPOTLIGHTS
        if ((LightShadowType & LightShadowType_SpotLight) == LightShadowType_SpotLight)
        {
            Attenuation *= SpotAttenuation(L, -SpotLightDirectionAndSpecularScale.xyz, SpotLightAnglesAndSoftTransitionScaleAndLightShadowType.xy) * FadedShadow;
        }
#endif

        // 累加光照结果.
#if !FULLY_ROUGH
        FMobileDirectLighting Lighting = MobileIntegrateBxDF(ShadingModelContext, GBuffer, PointNoL, MaterialParameters.CameraVector, PointH, PointNoH);
        Color += min(65000.0, (Attenuation) * LightColorAndFalloffExponent.rgb * (1.0 / PI) * (Lighting.Diffuse + Lighting.Specular * SpotLightDirectionAndSpecularScale.w));
#else
        Color += (Attenuation * PointNoL) * LightColorAndFalloffExponent.rgb * (1.0 / PI) * ShadingModelContext.DiffuseColor;
#endif
    }
}
#endif

(......)

// 计算非直接光照.
half ComputeIndirect(VTPageTableResult LightmapVTPageTableResult, FVertexFactoryInterpolantsVSToPS Interpolants, float3 DiffuseDir, FMobileShadingModelContext ShadingModelContext, out half IndirectIrradiance, out half3 Color)
{
    //To keep IndirectLightingCache conherence with PC, initialize the IndirectIrradiance to zero.
    IndirectIrradiance = 0;
    Color = 0;

    // 非直接漫反射.
#if LQ_TEXTURE_LIGHTMAP
    float2 LightmapUV0, LightmapUV1;
    uint LightmapDataIndex;
    GetLightMapCoordinates(Interpolants, LightmapUV0, LightmapUV1, LightmapDataIndex);

    half4 LightmapColor = GetLightMapColorLQ(LightmapVTPageTableResult, LightmapUV0, LightmapUV1, LightmapDataIndex, DiffuseDir);
    Color += LightmapColor.rgb * ShadingModelContext.DiffuseColor * View.IndirectLightingColorScale;
    IndirectIrradiance = LightmapColor.a;
#elif CACHED_POINT_INDIRECT_LIGHTING
    #if MATERIALBLENDING_MASKED || MATERIALBLENDING_SOLID
        // 将法线应用到半透明物体.
        FThreeBandSHVectorRGB PointIndirectLighting;
        PointIndirectLighting.R.V0 = IndirectLightingCache.IndirectLightingSHCoefficients0[0];
        PointIndirectLighting.R.V1 = IndirectLightingCache.IndirectLightingSHCoefficients1[0];
        PointIndirectLighting.R.V2 = IndirectLightingCache.IndirectLightingSHCoefficients2[0];

        PointIndirectLighting.G.V0 = IndirectLightingCache.IndirectLightingSHCoefficients0[1];
        PointIndirectLighting.G.V1 = IndirectLightingCache.IndirectLightingSHCoefficients1[1];
        PointIndirectLighting.G.V2 = IndirectLightingCache.IndirectLightingSHCoefficients2[1];

        PointIndirectLighting.B.V0 = IndirectLightingCache.IndirectLightingSHCoefficients0[2];
        PointIndirectLighting.B.V1 = IndirectLightingCache.IndirectLightingSHCoefficients1[2];
        PointIndirectLighting.B.V2 = IndirectLightingCache.IndirectLightingSHCoefficients2[2];

        FThreeBandSHVector DiffuseTransferSH = CalcDiffuseTransferSH3(DiffuseDir, 1);

        // 计算加入了法线影响的漫反射光照.
        half3 DiffuseGI = max(half3(0, 0, 0), DotSH3(PointIndirectLighting, DiffuseTransferSH));

        IndirectIrradiance = Luminance(DiffuseGI);
        Color += ShadingModelContext.DiffuseColor * DiffuseGI * View.IndirectLightingColorScale;
    #else
        // 半透明使用无方向(Non-directional), 漫反射被打包在xyz, 已经在cpu端除了PI和SH漫反射.
        half3 PointIndirectLighting = IndirectLightingCache.IndirectLightingSHSingleCoefficient.rgb;
        half3 DiffuseGI = PointIndirectLighting;

        IndirectIrradiance = Luminance(DiffuseGI);
        Color += ShadingModelContext.DiffuseColor * DiffuseGI * View.IndirectLightingColorScale;
    #endif
#endif

    return IndirectIrradiance;
}

// PS主入口.
PIXELSHADER_EARLYDEPTHSTENCIL
void Main(
    FVertexFactoryInterpolantsVSToPS Interpolants
    , FMobileBasePassInterpolantsVSToPS BasePassInterpolants
    , in float4 SvPosition : SV_Position
    OPTIONAL_IsFrontFace
    , out half4 OutColor    : SV_Target0
#if DEFERRED_SHADING_PATH
    , out half4 OutGBufferA    : SV_Target1
    , out half4 OutGBufferB    : SV_Target2
    , out half4 OutGBufferC    : SV_Target3
#endif
#if USE_SCENE_DEPTH_AUX
    , out float OutSceneDepthAux : SV_Target4
#endif
#if OUTPUT_PIXEL_DEPTH_OFFSET
    , out float OutDepth : SV_Depth
#endif
    )
{
#if MOBILE_MULTI_VIEW
    ResolvedView = ResolveView(uint(BasePassInterpolants.MultiViewId));
#else
    ResolvedView = ResolveView();
#endif

#if USE_PS_CLIP_PLANE
    clip(BasePassInterpolants.OutClipDistance);
#endif

    // 解压打包的插值数据.
#if PACK_INTERPOLANTS
    float4 PackedInterpolants[NUM_VF_PACKED_INTERPOLANTS];
    VertexFactoryUnpackInterpolants(Interpolants, PackedInterpolants);
#endif

#if COMPILER_GLSL_ES3_1 && !OUTPUT_MOBILE_HDR && !MOBILE_EMULATION
    // LDR Mobile needs screen vertical flipped
    SvPosition.y = ResolvedView.BufferSizeAndInvSize.y - SvPosition.y - 1;
#endif

    // 获取材质的像素属性.
    FMaterialPixelParameters MaterialParameters = GetMaterialPixelParameters(Interpolants, SvPosition);
    FPixelMaterialInputs PixelMaterialInputs;
    {
        float4 ScreenPosition = SvPositionToResolvedScreenPosition(SvPosition);
        float3 WorldPosition = BasePassInterpolants.PixelPosition.xyz;
        float3 WorldPositionExcludingWPO = BasePassInterpolants.PixelPosition.xyz;
        #if USE_WORLD_POSITION_EXCLUDING_SHADER_OFFSETS
            WorldPositionExcludingWPO = BasePassInterpolants.PixelPositionExcludingWPO;
        #endif
        CalcMaterialParametersEx(MaterialParameters, PixelMaterialInputs, SvPosition, ScreenPosition, bIsFrontFace, WorldPosition, WorldPositionExcludingWPO);

#if FORCE_VERTEX_NORMAL
        // Quality level override of material's normal calculation, can be used to avoid normal map reads etc.
        MaterialParameters.WorldNormal = MaterialParameters.TangentToWorld[2];
        MaterialParameters.ReflectionVector = ReflectionAboutCustomWorldNormal(MaterialParameters, MaterialParameters.WorldNormal, false);
#endif
    }

    // 像素深度偏移.
#if OUTPUT_PIXEL_DEPTH_OFFSET
    ApplyPixelDepthOffsetForMobileBasePass(MaterialParameters, PixelMaterialInputs, OutDepth);
#endif

    // Mask材质.
#if !EARLY_Z_PASS_ONLY_MATERIAL_MASKING
    //Clip if the blend mode requires it.
    GetMaterialCoverageAndClipping(MaterialParameters, PixelMaterialInputs);
#endif

    // 计算并缓存GBuffer数据, 防止后续多次采用纹理.
    FGBufferData GBuffer = (FGBufferData)0;
    GBuffer.WorldNormal = MaterialParameters.WorldNormal;
    GBuffer.BaseColor = GetMaterialBaseColor(PixelMaterialInputs);
    GBuffer.Metallic = GetMaterialMetallic(PixelMaterialInputs);
    GBuffer.Specular = GetMaterialSpecular(PixelMaterialInputs);
    GBuffer.Roughness = GetMaterialRoughness(PixelMaterialInputs);
    GBuffer.ShadingModelID = GetMaterialShadingModel(PixelMaterialInputs);
    half MaterialAO = GetMaterialAmbientOcclusion(PixelMaterialInputs);

    // 应用AO.
#if APPLY_AO
    half4 GatheredAmbientOcclusion = Texture2DSample(AmbientOcclusionTexture, AmbientOcclusionSampler, SvPositionToBufferUV(SvPosition));

    MaterialAO *= GatheredAmbientOcclusion.r;
#endif

    GBuffer.GBufferAO = MaterialAO;

    // 由于IEEE 754 (FP16)可表示的最小标准值是2^-24 = 5.96e-8, 而后面的粗糙度涉及到1.0 / Roughness^4的计算, 所以为了防止除零错误, 需保证Roughness^4 >= 5.96e-8, 此处直接Clamp粗糙度到0.015625(0.015625^4 = 5.96e-8).
    // 另外, 为了匹配PC端的延迟渲染(粗糙度存储在8位的值), 因此也自动Clamp到1.0.
    GBuffer.Roughness = max(0.015625, GetMaterialRoughness(PixelMaterialInputs));

    // 初始化移动端着色模型上下文FMobileShadingModelContext.
    FMobileShadingModelContext ShadingModelContext = (FMobileShadingModelContext)0;
    ShadingModelContext.Opacity = GetMaterialOpacity(PixelMaterialInputs);

    // 薄层透明度物
#if MATERIAL_SHADINGMODEL_THIN_TRANSLUCENT
    (......)
#endif

    half3 Color = 0;

    // 自定义数据.
    half CustomData0 = GetMaterialCustomData0(MaterialParameters);
    half CustomData1 = GetMaterialCustomData1(MaterialParameters);
    InitShadingModelContext(ShadingModelContext, GBuffer, MaterialParameters.SvPosition, MaterialParameters.CameraVector, CustomData0, CustomData1);
    float3 DiffuseDir = MaterialParameters.WorldNormal;

    // 头发模型.
#if MATERIAL_SHADINGMODEL_HAIR
    (......)
#endif

    // 光照图虚拟纹理.
    VTPageTableResult LightmapVTPageTableResult = (VTPageTableResult)0.0f;
#if LIGHTMAP_VT_ENABLED
    {
        float2 LightmapUV0, LightmapUV1;
        uint LightmapDataIndex;
        GetLightMapCoordinates(Interpolants, LightmapUV0, LightmapUV1, LightmapDataIndex);
        LightmapVTPageTableResult = LightmapGetVTSampleInfo(LightmapUV0, LightmapDataIndex, SvPosition.xy);
    }
#endif

#if LIGHTMAP_VT_ENABLED
    // This must occur after CalcMaterialParameters(), which is required to initialize the VT feedback mechanism
    // Lightmap request is always the first VT sample in the shader
    StoreVirtualTextureFeedback(MaterialParameters.VirtualTextureFeedback, 0, LightmapVTPageTableResult.PackedRequest);
#endif

    // 计算非直接光.
    half IndirectIrradiance;
    half3 IndirectColor;
    ComputeIndirect(LightmapVTPageTableResult, Interpolants, DiffuseDir, ShadingModelContext, IndirectIrradiance, IndirectColor);
    Color += IndirectColor;

    // 预计算的阴影图.
    half Shadow = GetPrimaryPrecomputedShadowMask(LightmapVTPageTableResult, Interpolants).r;

#if DEFERRED_SHADING_PATH
    float4 OutGBufferD;
    float4 OutGBufferE;
    float4 OutGBufferF;
    float4 OutGBufferVelocity = 0;

    GBuffer.IndirectIrradiance = IndirectIrradiance;
    GBuffer.PrecomputedShadowFactors.r = Shadow;

    // 编码GBuffer数据.
    EncodeGBuffer(GBuffer, OutGBufferA, OutGBufferB, OutGBufferC, OutGBufferD, OutGBufferE, OutGBufferF, OutGBufferVelocity);
#else

#if !MATERIAL_SHADINGMODEL_UNLIT

    // 天光.
#if ENABLE_SKY_LIGHT
    half3 SkyDiffuseLighting = GetSkySHDiffuseSimple(MaterialParameters.WorldNormal);
    half3 DiffuseLookup = SkyDiffuseLighting * ResolvedView.SkyLightColor.rgb;
    IndirectIrradiance += Luminance(DiffuseLookup);
#endif

    Color *= MaterialAO;
    IndirectIrradiance *= MaterialAO;

    float  ShadowPositionZ = 0;
#if DIRECTIONAL_LIGHT_CSM && !MATERIAL_SHADINGMODEL_SINGLELAYERWATER
    // CSM阴影.
    if (UseCSM())
    {
        half ShadowMap = MobileDirectionalLightCSM(MaterialParameters.ScreenPosition.xy, MaterialParameters.ScreenPosition.w, ShadowPositionZ);
    #if ALLOW_STATIC_LIGHTING
        Shadow = min(ShadowMap, Shadow);
    #else
        Shadow = ShadowMap;
    #endif
    }
#endif /* DIRECTIONAL_LIGHT_CSM */

    // 距离场阴影.
#if APPLY_DISTANCE_FIELD
    if (ShadowPositionZ == 0)
    {
        Shadow = Texture2DSample(MobileBasePass.ScreenSpaceShadowMaskTexture, MobileBasePass.ScreenSpaceShadowMaskSampler, SvPositionToBufferUV(SvPosition)).x;
    }
#endif

    half NoL = max(0, dot(MaterialParameters.WorldNormal, MobileDirectionalLight.DirectionalLightDirectionAndShadowTransition.xyz));
    half3 H = normalize(MaterialParameters.CameraVector + MobileDirectionalLight.DirectionalLightDirectionAndShadowTransition.xyz);
    half NoH = max(0, dot(MaterialParameters.WorldNormal, H));

    // 平行光 + IBL
#if FULLY_ROUGH
    Color += (Shadow * NoL) * MobileDirectionalLight.DirectionalLightColor.rgb * ShadingModelContext.DiffuseColor;
#else
    FMobileDirectLighting Lighting = MobileIntegrateBxDF(ShadingModelContext, GBuffer, NoL, MaterialParameters.CameraVector, H, NoH);
    // MobileDirectionalLight.DirectionalLightDistanceFadeMADAndSpecularScale.z保存了平行光的SpecularScale.
    Color += (Shadow) * MobileDirectionalLight.DirectionalLightColor.rgb * (Lighting.Diffuse + Lighting.Specular * MobileDirectionalLight.DirectionalLightDistanceFadeMADAndSpecularScale.z);

    // 头发着色.
#if    !(MATERIAL_SINGLE_SHADINGMODEL && MATERIAL_SHADINGMODEL_HAIR)
    (......)
#endif
#endif /* FULLY_ROUGH */

    // 局部光源, 最多4个.
#if MAX_DYNAMIC_POINT_LIGHTS > 0 && !MATERIAL_SHADINGMODEL_SINGLELAYERWATER

        if(NumDynamicPointLights > 0)
        {

            #if SUPPORT_SPOTLIGHTS_SHADOW
                FPCFSamplerSettings Settings;
                Settings.ShadowDepthTexture = DynamicSpotLightShadowTexture;
                Settings.ShadowDepthTextureSampler = DynamicSpotLightShadowSampler;
                Settings.ShadowBufferSize = DynamicSpotLightShadowBufferSize;
                Settings.bSubsurface = false;
                Settings.bTreatMaxDepthUnshadowed = false;
                Settings.DensityMulConstant = 0;
                Settings.ProjectionDepthBiasParameters = 0;
            #endif

            AccumulateLightingOfDynamicPointLight(MaterialParameters, ...);

            if (MAX_DYNAMIC_POINT_LIGHTS > 1 && NumDynamicPointLights > 1)
            {
                AccumulateLightingOfDynamicPointLight(MaterialParameters, ...);

                if (MAX_DYNAMIC_POINT_LIGHTS > 2 && NumDynamicPointLights > 2)
                {
                    AccumulateLightingOfDynamicPointLight(MaterialParameters, ...);

                    if (MAX_DYNAMIC_POINT_LIGHTS > 3 && NumDynamicPointLights > 3)
                    {
                        AccumulateLightingOfDynamicPointLight(MaterialParameters, ...);
                    }
                }
            }
        }

#endif

    // 天空光.
#if ENABLE_SKY_LIGHT
    #if MATERIAL_TWOSIDED && LQ_TEXTURE_LIGHTMAP
    if (NoL == 0)
    {
    #endif

    #if MATERIAL_SHADINGMODEL_SINGLELAYERWATER
        ShadingModelContext.WaterDiffuseIndirectLuminance += SkyDiffuseLighting;
    #endif
        Color += SkyDiffuseLighting * half3(ResolvedView.SkyLightColor.rgb) * ShadingModelContext.DiffuseColor * MaterialAO;
    #if MATERIAL_TWOSIDED && LQ_TEXTURE_LIGHTMAP
    }
    #endif
#endif

#endif /* !MATERIAL_SHADINGMODEL_UNLIT */

#if MATERIAL_SHADINGMODEL_SINGLELAYERWATER
    (......)
#endif // MATERIAL_SHADINGMODEL_SINGLELAYERWATER

#endif// DEFERRED_SHADING_PATH

    // 处理顶点雾.
    half4 VertexFog = half4(0, 0, 0, 1);
#if USE_VERTEX_FOG
#if PACK_INTERPOLANTS
    VertexFog = PackedInterpolants[0];
#else
    VertexFog = BasePassInterpolants.VertexFog;
#endif
#endif

    // 自发光.
    half3 Emissive = GetMaterialEmissive(PixelMaterialInputs);
#if MATERIAL_SHADINGMODEL_THIN_TRANSLUCENT
    Emissive *= TopMaterialCoverage;
#endif
    Color += Emissive;

#if !MATERIAL_SHADINGMODEL_UNLIT && MOBILE_EMULATION
    Color = lerp(Color, ShadingModelContext.DiffuseColor, ResolvedView.UnlitViewmodeMask);
#endif

    // 组合雾颜色到输出颜色.
    #if MATERIALBLENDING_ALPHACOMPOSITE || MATERIAL_SHADINGMODEL_SINGLELAYERWATER
        OutColor = half4(Color * VertexFog.a + VertexFog.rgb * ShadingModelContext.Opacity, ShadingModelContext.Opacity);
    #elif MATERIALBLENDING_ALPHAHOLDOUT
        // not implemented for holdout
        OutColor = half4(Color * VertexFog.a + VertexFog.rgb * ShadingModelContext.Opacity, ShadingModelContext.Opacity);
    #elif MATERIALBLENDING_TRANSLUCENT
        OutColor = half4(Color * VertexFog.a + VertexFog.rgb, ShadingModelContext.Opacity);
    #elif MATERIALBLENDING_ADDITIVE
        OutColor = half4(Color * (VertexFog.a * ShadingModelContext.Opacity.x), 0.0f);
    #elif MATERIALBLENDING_MODULATE
        half3 FoggedColor = lerp(half3(1, 1, 1), Color, VertexFog.aaa * VertexFog.aaa);
        OutColor = half4(FoggedColor, ShadingModelContext.Opacity);
    #else
        OutColor.rgb = Color * VertexFog.a + VertexFog.rgb;

        #if !MATERIAL_USE_ALPHA_TO_COVERAGE
            // Scene color alpha is not used yet so we set it to 1
            OutColor.a = 1.0;

            #if OUTPUT_MOBILE_HDR
                // Store depth in FP16 alpha. This depth value can be fetched during translucency or sampled in post-processing
                OutColor.a = SvPosition.z;
            #endif
        #else
            half MaterialOpacityMask = GetMaterialMaskInputRaw(PixelMaterialInputs);
            OutColor.a = GetMaterialMask(PixelMaterialInputs) / max(abs(ddx(MaterialOpacityMask)) + abs(ddy(MaterialOpacityMask)), 0.0001f) + 0.5f;
        #endif
    #endif

    #if !MATERIALBLENDING_MODULATE && USE_PREEXPOSURE
        OutColor.rgb *= ResolvedView.PreExposure;
    #endif

    #if MATERIAL_IS_SKY
        OutColor.rgb = min(OutColor.rgb, Max10BitsFloat.xxx * 0.5f);
    #endif

#if USE_SCENE_DEPTH_AUX
    OutSceneDepthAux = SvPosition.z;
#endif

    // 处理颜色的alpha.
#if USE_EDITOR_COMPOSITING && (MOBILE_EMULATION)
    // Editor primitive depth testing
    OutColor.a = 1.0;
    #if MATERIALBLENDING_MASKED
        // some material might have an opacity value
        OutColor.a = GetMaterialMaskInputRaw(PixelMaterialInputs);
    #endif
    clip(OutColor.a - GetMaterialOpacityMaskClipValue());
#else
    #if OUTPUT_GAMMA_SPACE
        OutColor.rgb = sqrt(OutColor.rgb);
    #endif
#endif

#if NUM_VIRTUALTEXTURE_SAMPLES || LIGHTMAP_VT_ENABLED
    FinalizeVirtualTextureFeedback(
        MaterialParameters.VirtualTextureFeedback,
        MaterialParameters.SvPosition,
        ShadingModelContext.Opacity,
        View.FrameNumber,
        View.VTFeedbackBuffer
    );
#endif
}

移动端的BasePassPS的处理过程比较复杂,步骤繁多,主要有:解压插值数据,获取并计算材质属性,计算并缓村GBuffer,处理或调整GBuffer数据,计算前向渲染分支的光照(平行光、局部光),计算距离场、CSM等阴影,计算天空光,处理静态光照、非直接光和IBL,计算雾效,以及处理水体、头发、薄层透明度等特殊着色模型。

由于标准16位浮点数(FP16)可表示的最小值是\(\cfrac{1.0}{2^{24}} = 5.96 \cdot 10^{-8}\),而后续的光照计算涉及到粗糙度的4次方运算(\(\cfrac{1.0}{\text{Roughness}^4}\)),为了防止除零错误,需要将粗糙度截取到\(0.015625\)(\(0.015625^4 = 5.96 \cdot 10^{-8}\))。

GBuffer.Roughness = max(0.015625, GetMaterialRoughness(PixelMaterialInputs));

这也警示我们在开发移动端的渲染特性时,需要格外注意和把控数据精度,否则在低端设备经常由于数据精度不足而出现各种奇葩的画面异常。

虽然上面的代码较多,但由很多宏控制着,实际渲染单个材质所需的代码可能只是其中的很小的一个子集。比如说,默认支持4个局部光源,但如果在工程配置(下图)中可以设为2或更少,则实际执行的光源指令少了很多。

如果是前向渲染分支,则GBuffer的很多处理将被忽略;如果是延迟渲染分支,则平行光、局部光源的计算将被忽略,由延迟渲染Pass的shader执行。

下面对重要接口EncodeGBuffer做剖析:

void EncodeGBuffer(
    FGBufferData GBuffer,
    out float4 OutGBufferA,
    out float4 OutGBufferB,
    out float4 OutGBufferC,
    out float4 OutGBufferD,
    out float4 OutGBufferE,
    out float4 OutGBufferVelocity,
    float QuantizationBias = 0        // -0.5 to 0.5 random float. Used to bias quantization.
    )
{
    if (GBuffer.ShadingModelID == SHADINGMODELID_UNLIT)
    {
        OutGBufferA = 0;
        SetGBufferForUnlit(OutGBufferB);
        OutGBufferC = 0;
        OutGBufferD = 0;
        OutGBufferE = 0;
    }
    else
    {
        // GBufferA: 八面体压缩后的法线, 预计算阴影因子, 逐物体数据.
#if MOBILE_DEFERRED_SHADING
        OutGBufferA.rg = UnitVectorToOctahedron( normalize(GBuffer.WorldNormal) ) * 0.5f + 0.5f;
        OutGBufferA.b = GBuffer.PrecomputedShadowFactors.x;
        OutGBufferA.a = GBuffer.PerObjectGBufferData;
#else
        (......)
#endif

        // GBufferB: 金属度, 高光度, 粗糙度, 着色模型, 其它Mask.
        OutGBufferB.r = GBuffer.Metallic;
        OutGBufferB.g = GBuffer.Specular;
        OutGBufferB.b = GBuffer.Roughness;
        OutGBufferB.a = EncodeShadingModelIdAndSelectiveOutputMask(GBuffer.ShadingModelID, GBuffer.SelectiveOutputMask);

        // GBufferC: 基础色, AO或非直接光.
        OutGBufferC.rgb = EncodeBaseColor( GBuffer.BaseColor );

#if ALLOW_STATIC_LIGHTING
        // No space for AO. Multiply IndirectIrradiance by AO instead of storing.
        OutGBufferC.a = EncodeIndirectIrradiance(GBuffer.IndirectIrradiance * GBuffer.GBufferAO) + QuantizationBias * (1.0 / 255.0);
#else
        OutGBufferC.a = GBuffer.GBufferAO;
#endif

        OutGBufferD = GBuffer.CustomData;
        OutGBufferE = GBuffer.PrecomputedShadowFactors;
    }

#if WRITES_VELOCITY_TO_GBUFFER
    OutGBufferVelocity = GBuffer.Velocity;
#else
    OutGBufferVelocity = 0;
#endif
}

在默认光照模型(DefaultLit)下,BasePass输出的结果有以下几种纹理:

12.3.3.3 MobileDeferredShading

移动端延迟光照的VS和PC端是一样的,都是DeferredLightVertexShaders.usf,但PS不一样,用的是MobileDeferredShading.usf。由于VS和PC一样,且没有特殊的操作,此处就忽略,如果有兴趣的同学可以看第五篇的小节5.5.3.1 DeferredLightVertexShader

下面直接分析PS代码:

// Engine\Shaders\Private\MobileDeferredShading.usf

(......)

// 移动端光源数据结构体.
struct FMobileLightData
{
    float3 Position;
    float  InvRadius;
    float3 Color;
    float  FalloffExponent;
    float3 Direction;
    float2 SpotAngles;
    float SourceRadius;
    float SpecularScale;
    bool bInverseSquared;
    bool bSpotLight;
};

// 获取GBuffer数据.
void FetchGBuffer(in float2 UV, out float4 GBufferA, out float4 GBufferB, out float4 GBufferC, out float4 GBufferD, out float SceneDepth)
{
    // Vulkan的子pass获取数据.
#if VULKAN_PROFILE
    GBufferA = VulkanSubpassFetch1();
    GBufferB = VulkanSubpassFetch2();
    GBufferC = VulkanSubpassFetch3();
    GBufferD = 0;
    SceneDepth = ConvertFromDeviceZ(VulkanSubpassDepthFetch());
    // Metal的子pass获取数据.
#elif METAL_PROFILE
    GBufferA = SubpassFetchRGBA_1();
    GBufferB = SubpassFetchRGBA_2();
    GBufferC = SubpassFetchRGBA_3();
    GBufferD = 0;
    SceneDepth = ConvertFromDeviceZ(SubpassFetchR_4());
    // 其它平台(DX, OpenGL)的子pass获取数据.
#else
    GBufferA = Texture2DSampleLevel(MobileSceneTextures.GBufferATexture, MobileSceneTextures.GBufferATextureSampler, UV, 0);
    GBufferB = Texture2DSampleLevel(MobileSceneTextures.GBufferBTexture, MobileSceneTextures.GBufferBTextureSampler, UV, 0);
    GBufferC = Texture2DSampleLevel(MobileSceneTextures.GBufferCTexture, MobileSceneTextures.GBufferCTextureSampler, UV, 0);
    GBufferD = 0;
    SceneDepth = ConvertFromDeviceZ(Texture2DSampleLevel(MobileSceneTextures.SceneDepthTexture, MobileSceneTextures.SceneDepthTextureSampler, UV, 0).r);
#endif
}

// 解压GBuffer数据.
FGBufferData DecodeGBufferMobile(
    float4 InGBufferA,
    float4 InGBufferB,
    float4 InGBufferC,
    float4 InGBufferD)
{
    FGBufferData GBuffer;
    GBuffer.WorldNormal = OctahedronToUnitVector( InGBufferA.xy * 2.0f - 1.0f );
    GBuffer.PrecomputedShadowFactors = InGBufferA.z;
    GBuffer.PerObjectGBufferData = InGBufferA.a;
    GBuffer.Metallic    = InGBufferB.r;
    GBuffer.Specular    = InGBufferB.g;
    GBuffer.Roughness    = max(0.015625, InGBufferB.b);
    // Note: must match GetShadingModelId standalone function logic
    // Also Note: SimpleElementPixelShader directly sets SV_Target2 ( GBufferB ) to indicate unlit.
    // An update there will be required if this layout changes.
    GBuffer.ShadingModelID = DecodeShadingModelId(InGBufferB.a);
    GBuffer.SelectiveOutputMask = DecodeSelectiveOutputMask(InGBufferB.a);
    GBuffer.BaseColor = DecodeBaseColor(InGBufferC.rgb);
#if ALLOW_STATIC_LIGHTING
    GBuffer.GBufferAO = 1;
    GBuffer.IndirectIrradiance = DecodeIndirectIrradiance(InGBufferC.a);
#else
    GBuffer.GBufferAO = InGBufferC.a;
    GBuffer.IndirectIrradiance = 1;
#endif
    GBuffer.CustomData = HasCustomGBufferData(GBuffer.ShadingModelID) ? InGBufferD : 0;
    return GBuffer;
}

// 直接光照.
half3 GetDirectLighting(
    FMobileLightData LightData,
    FMobileShadingModelContext ShadingModelContext,
    FGBufferData GBuffer,
    float3 WorldPosition,
    half3 CameraVector)
{
    half3 DirectLighting = 0;

    float3 ToLight = LightData.Position - WorldPosition;
    float DistanceSqr = dot(ToLight, ToLight);
    float3 L = ToLight * rsqrt(DistanceSqr);

    // 光源衰减.
    float Attenuation = 0.0;
    if (LightData.bInverseSquared)
    {
        // Sphere falloff (technically just 1/d2 but this avoids inf)
        Attenuation = 1.0f / (DistanceSqr + 1.0f);
        Attenuation *= Square(saturate(1 - Square(DistanceSqr * Square(LightData.InvRadius))));
    }
    else
    {
        Attenuation = RadialAttenuation(ToLight * LightData.InvRadius, LightData.FalloffExponent);
    }

    // 聚光灯衰减.
    if (LightData.bSpotLight)
    {
        Attenuation *= SpotAttenuation(L, -LightData.Direction, LightData.SpotAngles);
    }

    // 如果衰减不为0, 则计算直接光照.
    if (Attenuation > 0.0)
    {
        half3 H = normalize(CameraVector + L);
        half NoL = max(0.0, dot(GBuffer.WorldNormal, L));
        half NoH = max(0.0, dot(GBuffer.WorldNormal, H));
        FMobileDirectLighting Lighting = MobileIntegrateBxDF(ShadingModelContext, GBuffer, NoL, CameraVector, H, NoH);
        DirectLighting = (Lighting.Diffuse + Lighting.Specular * LightData.SpecularScale) * (LightData.Color * (1.0 / PI) * Attenuation);
    }
    return DirectLighting;
}

// 光照函数.
half ComputeLightFunctionMultiplier(float3 WorldPosition);
// 使用光网格添加局部光照, 不支持动态阴影, 因为需要逐光源阴影图.
half3 GetLightGridLocalLighting(const FCulledLightsGridData InLightGridData, ...);

// 平行光的PS主入口.
void MobileDirectLightPS(
    noperspective float4 UVAndScreenPos : TEXCOORD0,
    float4 SvPosition : SV_POSITION,
    out half4 OutColor : SV_Target0)
{
    // 恢复(读取)GBuffer数据.
    FGBufferData GBuffer = (FGBufferData)0;
    float SceneDepth = 0;
    {
        float4 GBufferA = 0;
        float4 GBufferB = 0;
        float4 GBufferC = 0;
        float4 GBufferD = 0;
        FetchGBuffer(UVAndScreenPos.xy, GBufferA, GBufferB, GBufferC, GBufferD, SceneDepth);
        GBuffer = DecodeGBufferMobile(GBufferA, GBufferB, GBufferC, GBufferD);
    }

    // 计算基础向量.
    float2 ScreenPos = UVAndScreenPos.zw;
    float3 WorldPosition = mul(float4(ScreenPos * SceneDepth, SceneDepth, 1), View.ScreenToWorld).xyz;
    half3 CameraVector = normalize(View.WorldCameraOrigin - WorldPosition);
    half NoV = max(0, dot(GBuffer.WorldNormal, CameraVector));
    half3 ReflectionVector = GBuffer.WorldNormal * (NoV * 2.0) - CameraVector;

    half3 Color = 0;
    // Check movable light param to determine if we should be using precomputed shadows
    half Shadow = LightFunctionParameters2.z > 0.0f ? 1.0f : GBuffer.PrecomputedShadowFactors.r;

    // CSM阴影.
#if APPLY_CSM
    float  ShadowPositionZ = 0;
    float4 ScreenPosition = SvPositionToScreenPosition(float4(SvPosition.xyz,SceneDepth));
    float ShadowMap = MobileDirectionalLightCSM(ScreenPosition.xy, SceneDepth, ShadowPositionZ);
    Shadow = min(ShadowMap, Shadow);
#endif

    // 着色模型上下文.
    FMobileShadingModelContext ShadingModelContext = (FMobileShadingModelContext)0;
    {
        half DielectricSpecular = 0.08 * GBuffer.Specular;
        ShadingModelContext.DiffuseColor = GBuffer.BaseColor - GBuffer.BaseColor * GBuffer.Metallic;    // 1 mad
        ShadingModelContext.SpecularColor = (DielectricSpecular - DielectricSpecular * GBuffer.Metallic) + GBuffer.BaseColor * GBuffer.Metallic;    // 2 mad
        // 计算环境的BRDF.
        ShadingModelContext.SpecularColor = GetEnvBRDF(ShadingModelContext.SpecularColor, GBuffer.Roughness, NoV);
    }

    // 局部光源.
    float2 LocalPosition = SvPosition.xy - View.ViewRectMin.xy;
    uint GridIndex = ComputeLightGridCellIndex(uint2(LocalPosition.x, LocalPosition.y), SceneDepth);
    // 分簇光源
#if USE_CLUSTERED
    {
        const uint EyeIndex = 0;
        const FCulledLightsGridData CulledLightGridData = GetCulledLightsGrid(GridIndex, EyeIndex);
        Color += GetLightGridLocalLighting(CulledLightGridData, ShadingModelContext, GBuffer, WorldPosition, CameraVector, EyeIndex, 0);
    }
#endif

    // 计算平行光.
    half NoL = max(0, dot(GBuffer.WorldNormal, MobileDirectionalLight.DirectionalLightDirectionAndShadowTransition.xyz));
    half3 H = normalize(CameraVector + MobileDirectionalLight.DirectionalLightDirectionAndShadowTransition.xyz);
    half NoH = max(0, dot(GBuffer.WorldNormal, H));
    FMobileDirectLighting Lighting;
    Lighting.Specular = ShadingModelContext.SpecularColor * CalcSpecular(GBuffer.Roughness, NoH);
    Lighting.Diffuse = ShadingModelContext.DiffuseColor;
    Color += (Shadow * NoL) * MobileDirectionalLight.DirectionalLightColor.rgb * (Lighting.Diffuse + Lighting.Specular * MobileDirectionalLight.DirectionalLightDistanceFadeMADAndSpecularScale.z);

    // 处理反射(IBL, 反射捕捉器).
#if APPLY_REFLECTION
    uint NumCulledEntryIndex = (ForwardLightData.NumGridCells + GridIndex) * NUM_CULLED_LIGHTS_GRID_STRIDE;
    uint NumLocalReflectionCaptures = min(ForwardLightData.NumCulledLightsGrid[NumCulledEntryIndex + 0], ForwardLightData.NumReflectionCaptures);
    uint DataStartIndex = ForwardLightData.NumCulledLightsGrid[NumCulledEntryIndex + 1];

    float3 SpecularIBL = CompositeReflectionCapturesAndSkylight(
        1.0f,
        WorldPosition,
        ReflectionVector,//RayDirection,
        GBuffer.Roughness,
        GBuffer.IndirectIrradiance,
        1.0f,
        0.0f,
        NumLocalReflectionCaptures,
        DataStartIndex,
        0,
        true);

    Color += SpecularIBL * ShadingModelContext.SpecularColor;
#elif APPLY_SKY_REFLECTION
    float SkyAverageBrightness = 1.0f;
    float3 SpecularIBL = GetSkyLightReflection(ReflectionVector, GBuffer.Roughness, SkyAverageBrightness);
    SpecularIBL *= ComputeMixingWeight(GBuffer.IndirectIrradiance, SkyAverageBrightness, GBuffer.Roughness);
    Color += SpecularIBL * ShadingModelContext.SpecularColor;
#endif
    // 天空光漫反射.
    half3 SkyDiffuseLighting = GetSkySHDiffuseSimple(GBuffer.WorldNormal);
    Color+= SkyDiffuseLighting * half3(View.SkyLightColor.rgb) * ShadingModelContext.DiffuseColor * GBuffer.GBufferAO;
    half LightAttenuation = ComputeLightFunctionMultiplier(WorldPosition);

#if USE_PREEXPOSURE
    // MobileHDR applies PreExposure in tonemapper
    LightAttenuation *= View.PreExposure;
#endif

    OutColor.rgb = Color.rgb * LightAttenuation;
    OutColor.a = 1;
}

// 局部光源的PS主入口.
void MobileRadialLightPS(
    float4 InScreenPosition : TEXCOORD0,
    float4 SVPos            : SV_POSITION,
    out half4 OutColor        : SV_Target0
)
{
    FGBufferData GBuffer = (FGBufferData)0;
    float SceneDepth = 0;
    {
        float2 ScreenUV = InScreenPosition.xy / InScreenPosition.w * View.ScreenPositionScaleBias.xy + View.ScreenPositionScaleBias.wz;
        float4 GBufferA = 0;
        float4 GBufferB = 0;
        float4 GBufferC = 0;
        float4 GBufferD = 0;
        FetchGBuffer(ScreenUV, GBufferA, GBufferB, GBufferC, GBufferD, SceneDepth);
        GBuffer = DecodeGBufferMobile(GBufferA, GBufferB, GBufferC, GBufferD);
    }

    // With a perspective projection, the clip space position is NDC * Clip.w
    // With an orthographic projection, clip space is the same as NDC
    float2 ClipPosition = InScreenPosition.xy / InScreenPosition.w * (View.ViewToClip[3][3] < 1.0f ? SceneDepth : 1.0f);
    float3 WorldPosition = mul(float4(ClipPosition, SceneDepth, 1), View.ScreenToWorld).xyz;
    half3 CameraVector = normalize(View.WorldCameraOrigin - WorldPosition);
    half NoV = max(0, dot(GBuffer.WorldNormal, CameraVector));

    // 组装光源数据结构体.
    FMobileLightData LightData = (FMobileLightData)0;
    {
        LightData.Position = DeferredLightUniforms.Position;
        LightData.InvRadius = DeferredLightUniforms.InvRadius;
        LightData.Color = DeferredLightUniforms.Color;
        LightData.FalloffExponent = DeferredLightUniforms.FalloffExponent;
        LightData.Direction = DeferredLightUniforms.Direction;
        LightData.SpotAngles = DeferredLightUniforms.SpotAngles;
        LightData.SpecularScale = 1.0;
        LightData.bInverseSquared = INVERSE_SQUARED_FALLOFF;
        LightData.bSpotLight = IS_SPOT_LIGHT;
    }

    FMobileShadingModelContext ShadingModelContext = (FMobileShadingModelContext)0;
    {
        half DielectricSpecular = 0.08 * GBuffer.Specular;
        ShadingModelContext.DiffuseColor = GBuffer.BaseColor - GBuffer.BaseColor * GBuffer.Metallic;    // 1 mad
        ShadingModelContext.SpecularColor = (DielectricSpecular - DielectricSpecular * GBuffer.Metallic) + GBuffer.BaseColor * GBuffer.Metallic;    // 2 mad
        // 计算环境BRDF.
        ShadingModelContext.SpecularColor = GetEnvBRDF(ShadingModelContext.SpecularColor, GBuffer.Roughness, NoV);
    }

    // 计算直接光.
    half3 Color = GetDirectLighting(LightData, ShadingModelContext, GBuffer, WorldPosition, CameraVector);

    // IES, 光照函数.
    half LightAttenuation = ComputeLightProfileMultiplier(WorldPosition, DeferredLightUniforms.Position, -DeferredLightUniforms.Direction, DeferredLightUniforms.Tangent);
    LightAttenuation*= ComputeLightFunctionMultiplier(WorldPosition);

#if USE_PREEXPOSURE
    // MobileHDR applies PreExposure in tonemapper
    LightAttenuation*= View.PreExposure;
#endif

    OutColor.rgb = Color * LightAttenuation;
    OutColor.a = 1;
}

以上可知,平行光和局部光源的PS是不同的入口,主要是因为两者的区别较大,平行光直接在主入口计算光照,附带计算了反射(IBL、捕捉器)、天空光漫反射;而局部光源会构建一个光源结构体,进入直接光计算函数,最后处理局部光源特有的IES和光照函数。

另外,获取GBuffer时,采用了SubPass特有的读取模式,不同的着色平台有所不同:

// Vulkan
[[vk::input_attachment_index(1)]]
SubpassInput<float4> GENERATED_SubpassFetchAttachment0;
#define VulkanSubpassFetch0() GENERATED_SubpassFetchAttachment0.SubpassLoad()

// Metal
Texture2D<float4> gl_LastFragDataRGBA_1;
#define SubpassFetchRGBA_1() gl_LastFragDataRGBA_1.Load(uint3(0, 0, 0), 0)

// DX / OpenGL
Texture2DSampleLevel(GBufferATexture, GBufferATextureSampler, UV, 0);

团队招员

博主所在的团队正在用UE4开发一种全新的沉浸式体验产品,急需各路豪士一同加入,共谋宏图大业。目前急招以下职位:

  • UE逻辑开发。
  • UE引擎程序。
  • UE图形渲染。
  • TA(技术向、美术向)。

要求:对技术有热情,扎实的技术基础,良好的沟通和合作能力,有UE使用经验或移动端开发经验者更佳。

有意向或想了解更多的请添加博主微信:81079389(注明博客园求职),或者发简历到博主邮箱:81079389#qq.com(#换成@)。

静待各路英雄豪杰相会。

特别说明

  • Part 1结束,Part 2的内容有:

    • 移动端渲染技术
    • 移动端优化技巧
  • 感谢所有参考文献的作者,部分图片来自参考文献和网络,侵删。

  • 本系列文章为笔者原创,只发表在博客园上,欢迎分享本文链接,但未经同意,不允许转载

  • 系列文章,未完待续,完整目录请戳内容纲目

  • 系列文章,未完待续,完整目录请戳内容纲目

  • 系列文章,未完待续,完整目录请戳内容纲目

参考文献

手机扫一扫

移动阅读更方便

阿里云服务器
腾讯云服务器
七牛云服务器

你可能感兴趣的文章