GameEngine on wanfeng

游戏引擎开发实践（降噪算法 SVGF）

Fri, 26 Dec 2025 15:26:32 +0800

降噪算法

高斯滤波器

很简单就是拿一个高斯滤波核遍历一遍像素

高斯公式回顾一下

双边滤波

这张图是高斯滤波的结果，它任何地方都会被均等的糊掉

下面是双边滤波器的公式 (i,j)(k,l)两个点的之间的权重计算，前半部分就是高斯分布，后半部分（也是一个高斯分布，颜色越相近，越接近最高点）表示：如果两个点颜色差距越大，权重越小

这个式子可以拆成两个指数函数的乘积，第二项如果颜色一致就=1，颜色不一致程度越大，越小，这样就达到了根据颜色差异调整权重的能力

这个算法的问题是分不清是边界还是噪声

联合双边滤波

高斯滤波的标准：距离

双边滤波的标准：距离+颜色差异

联合双边滤波思想就是用更多的指标，比如GBuffer那些信息

比如量化深度差异，法线差异来指导权重，这里并没有说具体公式

大滤波器优化

如果滤波核很大，如何进行优化

拆分2个Pass

水平一遍，竖直一遍，经典

再配合groupshared，减少ComputeShader的采样次数

原理上来讲二维高斯就可以拆成两个一维高斯的乘积，卷积写成二重积分后，对于高斯滤波核来说可以拆成内部x积分完成后，再y积分

但是复杂的滤波核不适用

小波变换卷积\空洞卷积 À-Trous wavelet

思路也是多个Pass进行卷积，比如5x5的核，每次都是5x5但是采样点的间距变了，变成$2^i$

复杂度从n2变成卷积核的平方 * 卷积次数

处理过亮像素

噪声中很亮的点如果做滤波，会变成一片亮

亮点探测:

处理亮点

SVGF （Spatiotemporal Variance-Guided Filter）时空方差引导滤波

SVGF就是融合了上边介绍的这些技术点，从时间和空间上都进行了滤波

3个因素引导的联合双边滤波

深度

它的深度指数衰减不仅仅是考虑空间深度，还会考虑法线投影上的深度大小（分母），这样建模是为了表达虽然空间上深度差异较大，但是xxxx(TODO:其实没懂为什么这样)

法线

两个法线的点乘来控制，并用指数控制衰减速度

另外一点不要用法线贴图的数据，直接用模型原始法线

Luminance

颜色差异会被噪声干扰，比如一个点在阴影里，但是因为噪声，它变得特别亮，解决办法就是利用方差来指导，方差信息会先计算一次空间上的方差，再时域累积，再空间滤波

SVGF解决不掉的问题是，场景没有变化也就是MotionVector没有数据，但是阴影一直在移动，这时候的阴影就会有噪声

再后来出现了A-SVGF,处理场景突变的延迟问题

RAE

利用神经网络来滤波，目前没兴趣

SVGF实践

参考文刀秋二大佬的https://zhuanlan.zhihu.com/p/28288053，与https://zhuanlan.zhihu.com/p/699706592

Demodulate Albedo

第一步要存储SPP=1的情况下，直接光照和间接光照的结果，并且在存储时要/albedo来保存，过滤过后再乘回来，文刀秋二的解释是（纹理的细节并不会因为Filter的强度过于大而丢失掉）目前不太理解，先做出来再说。另外这样做的前提应该是当前场景是Diffuse的

这张（右侧）表示直接光照/albedo

这张（右侧）表示间接光照/albedo

Reconstruction Filter

先实现空洞滤波

#version 450 core
#include "../common/common.glsl"
#include "../common/math.glsl"
#ifdef COMPUTE_SHADER
layout(set = 1, binding = 0, rgba32f) uniform image2D OUT_COLOR;
layout(set = 1, binding = 1, rgba32f) uniform image2D IN_COLOR;
layout(set = 1,binding = 2, rgba32f) uniform image2D velocityTexture;
layout(push_constant) uniform Uniforms
{
    int curPassIndex; // 当前 Atrous pass 索引
};

layout(local_size_x = 16, local_size_y = 16, local_size_z = 1) in;

void main(){
    ivec2 invocID = ivec2(gl_GlobalInvocationID.xy);
    ivec2 imgSize = ivec2(imageSize(OUT_COLOR));
    if (invocID.x >= imgSize.x || invocID.y >= imgSize.y) 
        return;

    vec3 colorSum = vec3(0.0);
    float weightSum = 0.0;


    // 5x5 Atrous 核心 做了空洞滤波，相当于64*64的卷积核
    int curStep = 1 << curPassIndex;
    for (int dy = -2; dy <= 2; ++dy) {
        for (int dx = -2; dx <= 2; ++dx) {
            ivec2 samplePos = invocID + ivec2(dx, dy) * curStep;

            // 边界检查
            samplePos = clamp(samplePos, ivec2(0), imgSize - 1);

            vec3 sampleColor = imageLoad(IN_COLOR, samplePos).xyz;

            // Atrous 权重，这里先用 1.0，之后可以换成高斯或 bilateral 权重
            float weight = 1.0;

            colorSum += sampleColor * weight;
            weightSum += weight;
        }
    }

    vec3 result = colorSum / weightSum;
    imageStore(OUT_COLOR, invocID, vec4(result, 1.0));
}
#endif

#include "hzpch.h"
#include "SVGFPass.h"
#include "Hazel/Core/Application.h"
#include "Hazel/Scene/SceneManager.h"
#include "Hazel/Renderer/RenderResource/RenderResourceManager.h"
#include "Hazel/Renderer/RenderResource/PipelineCache.h"
#include <Hazel/Renderer/RenderResource/Shader.h>
namespace GameEngine {
	void SVGFPass::Init()
	{
		m_Shader = std::make_shared<Shader>("postprocess/SVGF", SHADER_FREQUENCY_COMPUTE)->GetRHIShader();
		RHIRootSignatureInfo info = {};
		info.AddEntryFromReflect(m_Shader);
		info.AddEntry(RENDER_RESOURCEMANAGER->GetGlobalResourcePreFrameRootSignature()->GetInfo())
			.AddPushConstant({4,SHADER_FREQUENCY_COMPUTE });
		m_RootSignature = APP_DYNAMICRHI->CreateRootSignature(info);
		RHIComputePipelineInfo pipelineInfo = {};
		pipelineInfo.computeShader = m_Shader;
		pipelineInfo.rootSignature = m_RootSignature;
		m_Pipeline = APP_DYNAMICRHI->CreateComputePipeline(pipelineInfo);
	}

	void SVGFPass::Build(RDGBuilder& builder)
	{
		auto& [w, h] = APP_WINDOWSIZE;
		RDGTextureHandle pathTracingDirectRes = builder.GetTexture("PathTracingdirectRes");
		if (pathTracingDirectRes.ID() == UINT32_MAX) { return; }
		RDGTextureHandle pathTracingIndirectRes = builder.GetTexture("PathTracingIndirectRes");
		RDGTextureHandle velocity = builder.GetTexture("GBufferVelocity");
		RDGBufferHandle exposureData = builder.GetBuffer("ExposureData");

		RDGTextureHandle directFilterRes = builder.CreateTexture("PathTracing_SVGF_directFilterRes")
			.Exetent({ w, h ,1 })
			.Format(FORMAT_R32G32B32A32_SFLOAT)
			.AllowRenderTarget()
			.AllowReadWrite()
			.Finish();

		RDGTextureHandle forPinpongTexture = builder.CreateTexture("PathTracing_SVGF_forPinpongTexture")
			.Exetent({ w, h ,1 })
			.Format(FORMAT_R32G32B32A32_SFLOAT)
			.AllowRenderTarget()
			.AllowReadWrite()
			.Finish();


		RDGTextureHandle inDirectFilterRes = builder.CreateTexture("PathTracing_SVGF_inDirectFilterRes")
			.Exetent({ w, h ,1 })
			.Format(FORMAT_R32G32B32A32_SFLOAT)
			.AllowRenderTarget()
			.AllowReadWrite()
			.Finish();



		// 直接光
		for (int i = 0; i < 5; i++) {
			builder.CreateComputePass(GetName() + "_DirectRes" + std::to_string(i))
				.RootSignature(m_RootSignature)
				.PassIndex(i)
				.ReadWrite(1, 0, 0, i == 0 ? directFilterRes : i % 2 == 1 ? forPinpongTexture : directFilterRes)
				.ReadWrite(1, 1, 0, i == 0 ? pathTracingDirectRes : i % 2 == 1 ? directFilterRes : forPinpongTexture)
				.ReadWrite(1, 2, 0, velocity)
				.Execute([&](RDGPassContext context) {
				auto& [w, h] = APP_WINDOWSIZE;
				RHICommandListRef command = context.command;
				command->SetComputePipeline(m_Pipeline);
				command->BindDescriptorSet(context.descriptors[1], 1);
				uint32_t curIndex = context.passIndex[0];
				command->PushConstants(&curIndex, sizeof(uint32_t), SHADER_FREQUENCY_COMPUTE);
				command->BindDescriptorSet(RENDER_RESOURCEMANAGER->GetGlobalResourcePerFrameDescriptorSet(), 0);
				command->Dispatch((w + 15) / 16, (h + 15) / 16, 1);
					})
				.Finish();
		}



		// 间接光
		for (int i = 0; i < 5; i++) {
			builder.CreateComputePass(GetName() + "_DirectRes" + std::to_string(i))
				.RootSignature(m_RootSignature)
				.PassIndex(i)
				.ReadWrite(1, 0, 0, i == 0 ? inDirectFilterRes : i % 2 == 1 ? forPinpongTexture : inDirectFilterRes)
				.ReadWrite(1, 1, 0, i == 0 ? pathTracingIndirectRes : i % 2 == 1 ? inDirectFilterRes : forPinpongTexture)
				.ReadWrite(1, 2, 0, velocity)
				.Execute([&](RDGPassContext context) {
				auto& [w, h] = APP_WINDOWSIZE;
				RHICommandListRef command = context.command;
				command->SetComputePipeline(m_Pipeline);
				command->BindDescriptorSet(context.descriptors[1], 1);
				uint32_t curIndex = context.passIndex[0];
				command->PushConstants(&curIndex, sizeof(uint32_t), SHADER_FREQUENCY_COMPUTE);
				command->BindDescriptorSet(RENDER_RESOURCEMANAGER->GetGlobalResourcePerFrameDescriptorSet(), 0);
				command->Dispatch((w + 15) / 16, (h + 15) / 16, 1);
					})
				.Finish();
		}


	}

}

直接光模糊后结果

间接光会有闪烁的情况

滤波公式，每次滤波时都是一个联合双边滤波

具体的联合双边滤波有三部分组成

先加入法线贡献

// 5x5 Atrous 核心 做了空洞滤波，相当于64*64的卷积核
uint curStep = 1 << curPassIndex;
for (int dy = -2; dy <= 2; ++dy) {
    for (int dx = -2; dx <= 2; ++dx) {
        ivec2 samplePos = invocID + ivec2(dx, dy) * int(curStep);

        // 边界检查
        samplePos = clamp(samplePos, ivec2(0), imgSize - 1);

        vec3 sampleColor = imageLoad(IN_COLOR, samplePos).xyz;
        // 高斯
        float gaussWeight = gaussKernel5x5[dx + 2 + (dy + 2) * 5];
        // 法线
        vec3 sampleNormal = imageLoad(NORMAL, samplePos).xyz;
        float normalWeight = pow(max(dot(normal, sampleNormal), 0.0), 128.0);

        float weight = gaussWeight*normalWeight;

        colorSum += sampleColor * weight;
        weightSum += weight;
    }
}

明显还是有问题的，因为小正方形上面和地面法线都朝上，所以贡献偏大了，阴影跑到了正方形上边

需要引入颜色差异来解决这个问题，毕竟阴影里的颜色和阴影外的最大区别就是颜色，很容易区分，另外还需要注意颜色差异是噪声造成的还是场景造成的（方差越大，越不应该作为权重评判的指标）

分子表示亮度差，差异越大权重越小，但差异过大也可能是噪声导致的，所以差异大并不一定表示颜色边界

所以引入方差来表示颜色的变化，方差越大，权重越大。就是越不考虑颜色差异的影响

这里的方差是指时间维度的方差，这样的话就得用MotionVector来找上一帧的位置，也就可以理解为如果一个像素位置颜色一直在变化，那它就很有可能只是噪声，就不应该用颜色差异来知道滤波，所以这一项权重就会变大

先根据一阶矩和二阶矩计算出方差

时间上的方差结果只用于第一次滤波，后续4次滤波用一个公式来更新方差值

游戏引擎开发实践（DDGI）

Sun, 07 Dec 2025 20:36:06 +0800

总体流程

探针收集光照信息：探针向四面八方发射256条光线，计算击中点的光照并存储（此处可以额外计算击中点的间接光照，来模拟无限弹射的全局光照效果），另外还要存储hitT用来评估可见性

混合光照信息：每个探针分配6*6的像素，每个像素代表一个法线方向，通过蒙特卡洛积分来估计法线方向上的Irrandiance（其实和IBL思路一致），并进行时域加权混合，来获得更多的样本？

混合距离信息：每个探针分配 14*14的像素，计算每个方向上HitT的期望和平方的期望，时域混合（这两部都采用了八面体映射，另外这两步涉及到采样边界的问题，在像素外扩充一圈，用特定规则填充后存储）类似，保证在一个探针内采样时，不会采样到别的探针数据

光照Pass使用探针：根据着色点位置找到周围8个探针，从8个探针中获取对应法线位置的数据进行加权平均，有3种权重系数

如果probe离着色点较远，降低probe的权重（三线性插值系数）

如果着色点到probe的方向与表面法线的夹角过大，降低probe的权重（方向系数）

如果着色点与probe之间有较大的概率存在遮挡物，降低probe的权重（切比雪夫系数）

DDGI Probe

DDGIVolume

把探针组组成一个包围盒，定义探针数量和间距，以及DDGI的各种参数，这里实现了一个新的探针组件来表达它，并实现了他的可视化

ProbeTrace

这个阶段利用光追，计算并保存每个探针发射四面八方的光线的Radiance 信息（在每个击中点进行一次光照计算）

获取volume的设置，采用y-up方向，作为Image2Darray的layer，所以每层的探针数量为probeCount.x * probeCount.z

DDGISetting ddgiSetting = RENDER_RESOURCEMANAGER->GetGlobalSettingInfo().ddgiSetting;

glm::uvec3 probeCount = ddgiSetting.probeCount;
uint32_t raysPerProbe = ddgiSetting.raysPerProbe;
uint32_t probeCountPreLayer = probeCount.x * probeCount.z;
uint32_t volumeLayerCount = probeCount.y;  // y-up

执行光追TraceRays(raysPerProbe，probeCountPreLayer，volumeLayerCount)

// VolumeTrace
{
	auto& build = builder.CreateRayTracingPass(GetName() + "_VolumeTrace")
		.PassIndex(raysPerProbe, probeCountPreLayer, volumeLayerCount)
		.RootSignature(m_VolumeTraceRootSignature)
		.ReadWrite(1, 0, 0, rayTexture, VIEW_TYPE_2D_ARRAY, { TEXTURE_ASPECT_COLOR ,0,1,0,volumeLayerCount })
		.Read(1, 1, 0, dirShadowMap, VIEW_TYPE_2D_ARRAY, { TEXTURE_ASPECT_DEPTH ,0,1,0,4 })
		.Read(1, 3, 0, skyBox, VIEW_TYPE_CUBE, { TEXTURE_ASPECT_COLOR ,0,1,0,6 })
		.Read(1, 4, 0, irrandiance, VIEW_TYPE_2D_ARRAY, { TEXTURE_ASPECT_COLOR,0,1,0,volumeLayerCount })
		.Read(1, 5, 0, distance, VIEW_TYPE_2D_ARRAY, { TEXTURE_ASPECT_COLOR,0,1,0,volumeLayerCount })
		.Execute([&](RDGPassContext context) {
		RHICommandListRef command = context.command;
		command->SetRayTracingPipeline(m_VolumeTracePipeline);
		command->BindDescriptorSet(RENDER_RESOURCEMANAGER->GetGlobalResourcePerFrameDescriptorSet(), 0);
		command->BindDescriptorSet(context.descriptors[1], 1);
		command->TraceRays(context.passIndex[0], context.passIndex[1], context.passIndex[2]);
			});
	LightInfo& lightInfo = LightCollector::GetLightInfo();
	if (lightInfo.pointLightCount > 0) {
		RDGTextureHandle pointShadowMap = builder.GetTexture("Point Shadow Color[0]");
		build.Read(1, 2, 0, pointShadowMap, VIEW_TYPE_CUBE, { TEXTURE_ASPECT_COLOR,0,1,0,6 });
	}
}

光追Shader

RayGen中需要为每个探针的每条光线的击中点计算光照和阴影，并且混合历史信息（计算击中点的周围8个探针历史信息的irrandiance，作为击中点的间接光照（只计算漫反射部分））

struct Payload {
    vec3 albedo;
	float roughness; 
	vec3 worldPosition;
	float metallic;
	vec3 normal;
	float  hitT;
};


void main(){

    uint rayIndex = gl_LaunchIDEXT.x;
    uint probePlaneIndex = gl_LaunchIDEXT.y;
    uint planeIndex = gl_LaunchIDEXT.z;
    uint probeCountPrePlane = DDGIGetProbesPerPlane(volume.probeCount);  // x * z（y-up）
	uint probeIndex = (planeIndex * probeCountPrePlane) + probePlaneIndex; 
	uvec3 probeCoords = DDGIGetProbeCoords(probeIndex,volume); // 探针在探针网格的坐标
	vec3 probeWorldPosition = DDGIGetProbeWorldPosition(probeCoords, volume); 
	vec3 rayDirection = normalize(RTXGISphericalFibonacci(rayIndex,volume.raysPerProbe)); // 生成光线方向
    // 最终每根光线的计算结果都需要存储， Texture2DArray(x:rayIndex, y:probeIndexInLayer, z:layerIndex)
	uvec3 outputCoords = DDGIGetRayDataTexelCoords(rayIndex,probeIndex,volume);

    // 启动射线
	traceRayEXT(TLAS, 					// acceleration structure
		gl_RayFlagsOpaqueEXT,       	// rayFlags 控制光线的行为，比如是否忽略背面、是否启用 any-hit、是否可用 conservative tracing 等
		0xFF,           				// cullMask 0xFF → 匹配所有实例
		0,              				// sbtRecordOffset 索引到 Shader Binding Table (SBT) 的起始记录位置
		1,              				// sbtRecordStride SBT 中每条记录的间隔（单位是记录数，不是字节）
		0,              				// missIndex 当光线没有击中任何几何体时，使用 SBT 中 miss shader 的索引
		probeWorldPosition.xyz,     	// ray origin
		0,       						// ray min range
		rayDirection.xyz,  				// ray direction
		MAX_RAY_TRACING_DISTANCE,       // TODO:这个应该换成DDGI自己的参数设置
		0               				// payload (location = 0) payload的位置
  	);

    // Miss，返回天空的Radiance
    if(payload.hitT == -1.f){
        imageStore(out_RAYDATA, ivec3(outputCoords), vec4(payload.albedo, 1e27f));
        return;
	}

    // 击中背面，参考RTXGI的做法
    // Make the hit distance negative to mark a backface hit for blending, probe relocation, and probe classification.
	// Shorten the hit distance on a backface hit by 80% to decrease the influence of the probe during irradiance sampling.
    if(payload.hitKind == gl_HitKindBackFacingTriangleEXT){
    	imageStore(out_RAYDATA, ivec3(outputCoords), vec4(vec3(0), -payload.hitT * 0.2));
    	return;
	}

/////////////////////////////////////////////
// Directional Light TODO: 点光源和聚光   这里也是通过TraceRays来判断阴影的
/////////////////////////////////////////////
	vec3 brdf = (payload.albedo / PI);
	vec3 lighting = vec3(0.f);
	DirectionLight dirLight = GetDirectionLight();
	// 硬件在找到第一个 hit 后立即停止 | 跳过ClosestHitShader，直接返回rayGen | 所有模型都被视为不透明
	const uint rayFlags =gl_RayFlagsTerminateOnFirstHitEXT | gl_RayFlagsSkipClosestHitShaderEXT | gl_RayFlagsOpaqueEXT;
	// 发射光线计算遮挡
	traceRayEXT(TLAS,
		rayFlags,
		0xFF, 
		0, 
		1,
		1, 
		payload.worldPosition,
		0,   
		-dirLight.direction,
		MAX_RAY_TRACING_DISTANCE,
		1
  	);
	float shadowScale = 1.0;
	if(shadowPayload.hitT !=  -1.f){
		shadowScale = 0.0;
	}
    vec3 lightDirection = -normalize(dirLight.direction);
    float  nol = max(dot(payload.normal, lightDirection), 0.f);
	vec3 dirLighting = nol * dirLight.radiance * shadowScale;
	lighting += dirLighting;
	vec3 diffuse = lighting * brdf;


/////////////////////////////////////////////
// Indirection Light
/////////////////////////////////////////////
	vec3 irradiance = vec3(0);
	float volumeBlendWeight = DDGIGetVolumeBlendWeight(payload.worldPosition, volume);
	if (volumeBlendWeight > 0){

        irradiance = DDGIGetIrrandianceByWorldPosition(
            payload.worldPosition,
            payload.normal,
            volume,
            ddgi_Irrandiance,ddgi_Distance);
	}
	// Perfectly diffuse reflectors don't exist in the real world.
    // Limit the BRDF albedo to a maximum value to account for the energy loss at each bounce.
    float maxAlbedo = 0.9f;

    vec3 radiance = diffuse + ((min(payload.albedo, vec3(maxAlbedo, maxAlbedo, maxAlbedo)) / PI) * irradiance * volumeBlendWeight);

	// 最终存储rayData radiance(3) + hitT(1)
	imageStore(out_RAYDATA, ivec3(outputCoords), vec4(Saturate(radiance), payload.hitT));
}

closestHit需要获取击中点的模型信息

#ifdef RAYCLOSEST_HIT_SHADER
layout(location = 0) rayPayloadInEXT Payload payload;
hitAttributeEXT vec2 attribs;
void main()
{
	uint instanceID       = gl_InstanceCustomIndexEXT;   // 在构建TLAS时，给每个实例分配的ID
	uint primitiveID     = gl_PrimitiveID;  // 击中的三角形索引

	vec3 barycentrics = vec3(1.0 - attribs.x - attribs.y, attribs.x, attribs.y); 

	// 收集MeshInfo
	mat4 model = GetModelMatrix(instanceID);
	uvec3 triangleIndex = GetTriangleIndex(instanceID, primitiveID);
	vec4 position = GetTrianglePosition(instanceID,triangleIndex,barycentrics);
	vec3 meshNormal = GetTriangleMeshNormal(instanceID,triangleIndex,barycentrics);
	vec4 tangent = GetTriangleTangent(instanceID,triangleIndex,barycentrics);

	vec3 worldNormal = GetWorldNormal(meshNormal,model);
	vec4 worldTangent = GetWorldTangent(tangent,model);
	vec2 texCoord  = GetTriangleTexCoord(instanceID, triangleIndex, barycentrics);    
    vec4 worldPos       = model * position; 
	MaterialInfo material   = GetMaterialInfo(instanceID);
	vec4 albedo = GetDiffuse(material,texCoord);
	vec3 normal = GetNormal(material, texCoord, worldNormal, worldTangent);
    vec4 emission = GetEmission(material,texCoord);
	albedo += emission;
	float roughness = GetRoughness(material,texCoord);
    float metallic = GetMetallic(material, texCoord);

	payload.worldPosition = worldPos.xyz;
	payload.normal = normal;
	payload.roughness = roughness;
	payload.metallic = metallic;
	payload.albedo = albedo.rgb;
	payload.hitT = gl_HitTEXT;
	payload.hitKind = gl_HitKindEXT;
}
#endif

Miss处理未击中的情况，需要两个，一个处理天空，一个处理阴影判断标记（因为MissShader只支持一个Payload，好像只有RayGen可以定义多个PayLoad，其他都只能用于处理一个PayLoad）

#ifdef RAYMISS_SHADER
layout(location = 0) rayPayloadInEXT Payload payload;    // rayPayloadInEXT 注意是InEXT

layout(set = 1, binding = 3) uniform textureCube skyCube;
void main()
{
	vec3 rayDir = normalize(gl_WorldRayDirectionEXT);
    payload.albedo = texture(samplerCube(skyCube,SAMPLER[0]),rayDir).rgb;
	payload.hitT = -1;
}
#endif

#ifdef RAYMISS_SHADER
layout(location = 1) rayPayloadInEXT ShadowPayLoad shadowPayload;

void main()
{
	shadowPayload.hitT = -1;
}
#endif

RayTrace结束后，获得了每个探针的每条光线携带的Radiance信息

另外还会在着色点采样它周围8个探针的光照数据（就是上一帧的记录，也是该着色点的间接光照）这样做相当于得到了采样点的间接光照

ProbeBlend

根据RayData来更新探针信息

IrrandianceBlend

探针理论上需要存储每个方向上的半球积分来得到该位置的Irrandiance，但是存储所有方向不太可能，DDGI采用6*6的像素存储一个探针36个方向上的Irrandiance，使用时进行插值即可，另外使用八面体映射，遍历每个探针，保存一个6 * 6区域，每个区域都被映射到八面体上的一个方向，再把这个方向挪到球面，遍历所有光线与该方向的权重进行混合

注意Texture2DArray的采样的uvw的w不是0-1,用第几层就传递几，这点疏忽了给我整麻了

// dispatch 需要处理网格中的每一个探针，所以传入的是probeCount.x, probeCount.y, probeCount.z
{
        // Irrandiance Blend
        builder.CreateComputePass(GetName() + "_ProbeIrrandianceBlend")
            .ReadWrite(1, 0, 0, irrandiance, VIEW_TYPE_2D_ARRAY, { TEXTURE_ASPECT_COLOR ,0,1,0,volumeLayerCount })
            .ReadWrite(1, 1, 0, rayTexture, VIEW_TYPE_2D_ARRAY, { TEXTURE_ASPECT_COLOR ,0,1,0,volumeLayerCount })
            .RootSignature(m_ProbeIrrandianceBlendRootSignature)
            .PassIndex(probeCount.x, probeCount.z, probeCount.y)
            .Execute([&](RDGPassContext context) {
            RHICommandListRef command = context.command;
            command->SetComputePipeline(m_ProbeIrrandianceBlendPipeline);
            command->BindDescriptorSet(RENDER_RESOURCEMANAGER->GetGlobalResourcePerFrameDescriptorSet(), 0);
            command->BindDescriptorSet(context.descriptors[1], 1);
            command->Dispatch(context.passIndex[0], context.passIndex[1], context.passIndex[2]);
                })
            .Finish();
    }

layout(set = 1, rgba32f, binding = 0) uniform image2DArray o_Texture;
layout(set = 1, rgba32f, binding = 1) uniform image2DArray  IN_RayData;
// 每个探针需要8*8的区域，所以这里线程组设置就是8*8（内部6*6是数据，外部一圈是填充，为了保证双线性插值的正确性）
layout(local_size_x = DDGI_PROBE_NUM_TEXELS_IRRANDIANCE, local_size_y = DDGI_PROBE_NUM_TEXELS_IRRANDIANCE, local_size_z = 1) in;

void main(){
    // 索引探针在RayData的位置
    DDGISetting volume = GetDDGISetting();
    uvec3 groupID = gl_WorkGroupID;  // 对于dispatch的ID
    uvec3 invocationID = gl_GlobalInvocationID; // 相对于全局的调用ID 用于最后存储像素位置
    uvec3 LocalInvocationID = gl_LocalInvocationID; // 相对于组内的调用ID 用于区别边界和内部
    // 用LocalInvocationID是不是0，0或者 7，7 马上就能判断是不是边界
    bool isBorderTexel = (LocalInvocationID.x == 0 || LocalInvocationID.x == (DDGI_PROBE_NUM_TEXELS_IRRANDIANCE_INTERIOR + 1)) || (LocalInvocationID.y == 0 || LocalInvocationID.y == (DDGI_PROBE_NUM_TEXELS_IRRANDIANCE_INTERIOR + 1));
    
    // 
    uint probeIndex = (groupID.y * DDGIGetProbesPerPlane(volume.probeCount)) + groupID.z * volume.probeCount.x + groupID.x ;

    // Early out: no probe maps to this thread 感觉不会出现这种情况
    uint numProbes = (volume.probeCount.x * volume.probeCount.y * volume.probeCount.z);
    if (probeIndex >= numProbes || probeIndex < 0) return;
    
    vec4 result = vec4(0.0);
    
    // 首先处理内部6*6
    if(!isBorderTexel){
        // 对于6*6区域的一个像素，对应到探针的一个方向上，需要计算这个方向上的半球积分来存储Irrandiance
        // 1. 获取这个像素对应的方向
        uvec3 threadCoords = uvec3(groupID.x * DDGI_PROBE_NUM_TEXELS_IRRANDIANCE_INTERIOR, groupID.y * DDGI_PROBE_NUM_TEXELS_IRRANDIANCE_INTERIOR, invocationID.z) + LocalInvocationID - uvec3(1, 1, 0); // 计算出抛开边界后当前线程是的坐标（没有边界填充的2DArray坐标）
        // 把6*6区域当作一个纹理，下面获得当前线程处理的位置的UV坐标[-1,1]
        vec2 probeOctantUV = DDGIGetNormalizedOctahedralCoordinates(uvec2(threadCoords.xy), DDGI_PROBE_NUM_TEXELS_IRRANDIANCE_INTERIOR);
        // 获取这个UV对应到八面体上的方向
    	vec3 probeRayDirection = DDGIGetOctahedralDirection(probeOctantUV);

        
        // 遍历当前探针的所有光线，为当前八面体上的方向计算irrandiance
        for (uint rayIndex = 0; rayIndex < volume.raysPerProbe; rayIndex++){
            // 获取ray的方向，RTXGISphericalFibonacci是可以复现的所以不需要提前存储
            vec3 rayDirection = normalize(RTXGISphericalFibonacci(rayIndex,volume.raysPerProbe));  // TODO: 优化点，因为每个探针计算这个都是一模一样的
            // 当前要存储的方向与这跟光线的cos，两个方向夹角越小，权重越大
            float weight = max(0.0, dot(probeRayDirection, rayDirection));
            // 解析RayData中的信息
            uvec3 rayDataTexCoords = DDGIGetRayDataTexelCoords(rayIndex, probeIndex, volume);  
            vec4 storedData = imageLoad(IN_RayData, ivec3(rayDataTexCoords));
            vec3 probeRayRadiance = storedData.xyz;
            float probeRayDistance = storedData.w;
            if(probeRayDistance < 0){ // 击中的是背面
                continue;
            }
            result += vec4(probeRayRadiance * weight, weight);
        }

        // 到这里 用vec4存储了所有光线的加权和，并且w分量存储了权重和，方便计算
//////////////////////////////工程化修正/////////////////////////////////////////////
        float epsilon = float(volume.raysPerProbe);
        epsilon *= 1e-9f;
        result.rgb *= 1.f / (1.f * max(result.a, epsilon));

        // 时域加权混合
        vec4 history = imageLoad(o_Texture,ivec3(gl_GlobalInvocationID));
        vec3 delta = (result.rgb - history.rgb);

        float  hysteresis = 0.97; // TODO：混合系数是参数
        if (dot(history, history) == 0) hysteresis = 0.f;
        float probeIrradianceEncodingGamma = 5.0f;  // TODO：编码Gamma是参数
        
        result.rgb = pow(result.rgb, vec3(1.f / probeIrradianceEncodingGamma));
        vec3 irradianceSample = result.rgb;
        float probeIrradianceThreshold = 0.25f; // TODO：阈值是参数
        if (RTXGIMaxComponent(history.rgb - result.rgb) > probeIrradianceThreshold)
        {
            // Lower the hysteresis when a large lighting change is detected
            hysteresis = max(0.f, hysteresis - 0.75f);
        }
        float probeBrightnessThreshold  = 0.10f; // TODO：阈值是参数
        if (RGBtoLuminance(delta) > probeBrightnessThreshold)
        {
            // Clamp the maximum per-update change in irradiance when a large brightness change is detected
            delta *= 0.25f;
        }
        const float c_threshold = 1.f / 1024.f;
        vec3 lerpDelta = (1.f - hysteresis) * delta;
        if (RTXGIMaxComponent(result.rgb) < RTXGIMaxComponent(history.rgb))
        {
            lerpDelta = min(max(vec3(c_threshold), abs(lerpDelta)), abs(delta)) * sign(lerpDelta);
        }
        result = vec4(history.rgb + lerpDelta, 1.f);
        imageStore(o_Texture, ivec3(gl_GlobalInvocationID), result);
///////////////////////////////////////////////////////////////////////////
        
        
        // 边界处理
       else{
        // 边界
        // AllMemoryBarrierWithGroupSync();   这里有一个同步点 

        bool isCornerTexel = (LocalInvocationID.x == 0 || LocalInvocationID.x == (DDGI_PROBE_NUM_TEXELS_IRRANDIANCE - 1)) && (LocalInvocationID.y == 0 || LocalInvocationID.y == (DDGI_PROBE_NUM_TEXELS_IRRANDIANCE - 1));
        bool isRowTexel = (LocalInvocationID.x > 0 && LocalInvocationID.x < (DDGI_PROBE_NUM_TEXELS_IRRANDIANCE - 1));

        uvec3 copyCoordinates = uvec3(groupID.x * DDGI_PROBE_NUM_TEXELS_IRRANDIANCE, groupID.y * DDGI_PROBE_NUM_TEXELS_IRRANDIANCE, invocationID.z);

        if(isCornerTexel)
        {
            copyCoordinates.x += LocalInvocationID.x > 0 ? 1 : DDGI_PROBE_NUM_TEXELS_IRRANDIANCE_INTERIOR;
            copyCoordinates.y += LocalInvocationID.y > 0 ? 1 : DDGI_PROBE_NUM_TEXELS_IRRANDIANCE_INTERIOR;
        }
        else if(isRowTexel)
        {
            copyCoordinates.x += (DDGI_PROBE_NUM_TEXELS_IRRANDIANCE - 1) - LocalInvocationID.x;
            copyCoordinates.y += LocalInvocationID.y + ((LocalInvocationID.y > 0) ? -1 : 1);
        }
        else // Column Texel
        {
            copyCoordinates.x += LocalInvocationID.x + ((LocalInvocationID.x > 0) ? -1 : 1);
            copyCoordinates.y += (DDGI_PROBE_NUM_TEXELS_IRRANDIANCE - 1) - LocalInvocationID.y;
        }

        imageStore(o_Texture, ivec3(gl_GlobalInvocationID), imageLoad(o_Texture, ivec3(copyCoordinates)));
    }
}

DistanceBlend

对于distance，只有一些地方不一样

#version 450 core
#include "../common/common.glsl"
#include "../common/DDGI.glsl"
layout(set = 1, rgba32f, binding = 0) uniform image2DArray o_Texture;
layout(set = 1,rgba32f, binding = 1) uniform image2DArray  IN_RayData;
#ifdef COMPUTE_SHADER
layout(local_size_x = DDGI_PROBE_NUM_TEXELS_DISTANCE, local_size_y = DDGI_PROBE_NUM_TEXELS_DISTANCE, local_size_z = 1) in;  
void main(){
	DDGISetting volume = GetDDGISetting();
    uvec3 groupID = gl_WorkGroupID;  // 对于dispatch的ID
    uvec3 invocationID = gl_GlobalInvocationID; // 相对于全局的调用ID
    uvec3 LocalInvocationID = gl_LocalInvocationID; // 相对于组内的调用ID
    bool isBorderTexel = (LocalInvocationID.x == 0 || LocalInvocationID.x == (DDGI_PROBE_NUM_TEXELS_DISTANCE_INTERIOR + 1)) || (LocalInvocationID.y == 0 || LocalInvocationID.y == (DDGI_PROBE_NUM_TEXELS_DISTANCE_INTERIOR + 1));
    uint probeIndex = (groupID.y * DDGIGetProbesPerPlane(volume.probeCount)) + groupID.z * volume.probeCount.x + groupID.x ;

    vec4 result = vec4(0.0);

    if(!isBorderTexel){
        uvec3 threadCoords = uvec3(groupID.x * DDGI_PROBE_NUM_TEXELS_DISTANCE_INTERIOR, groupID.y * DDGI_PROBE_NUM_TEXELS_DISTANCE_INTERIOR, invocationID.z) + LocalInvocationID - uvec3(1, 1, 0);
        vec2 probeOctantUV = DDGIGetNormalizedOctahedralCoordinates(uvec2(threadCoords.xy), DDGI_PROBE_NUM_TEXELS_DISTANCE_INTERIOR);
        vec3 probeRayDirection = DDGIGetOctahedralDirection(probeOctantUV);
        // 遍历当前探针的所有光线
        for (uint rayIndex = 0; rayIndex < volume.raysPerProbe; rayIndex++){
            vec3 rayDirection = normalize(RTXGISphericalFibonacci(rayIndex, volume.raysPerProbe));
            float weight = max(0.f, dot(probeRayDirection, rayDirection));
            float probeDistanceExponent = 50.f; // TODO:参数
             // Increase or decrease the filtered distance value's "sharpness"
            weight = pow(weight, probeDistanceExponent);
            uvec3 rayDataTexCoords = DDGIGetRayDataTexelCoords(rayIndex, probeIndex, volume);
            // 光线最大能射到probe间距的1.5倍
            float probeMaxRayDistance = length(volume.gridStep) * 1.5f;
            vec4 storedData = imageLoad(IN_RayData, ivec3(rayDataTexCoords));
            float probeRayDistance = storedData.w;
            probeRayDistance = min(abs(probeRayDistance),probeMaxRayDistance);
            result += vec4(probeRayDistance * weight, (probeRayDistance * probeRayDistance) * weight, 0.f, weight);
        }
//////////////////////////////工程化修正/////////////////////////////////////////////
        float epsilon = float(volume.raysPerProbe);
        epsilon *= 1e-9f;
        result.rgb *= 1.f / (1.f * max(result.a, epsilon));

        // 时域加权混合
        vec4 history = imageLoad(o_Texture,ivec3(gl_GlobalInvocationID));
        float  hysteresis = 0.97; // TODO：混合系数是参数
        if (dot(history, history) == 0) hysteresis = 0.f;
        result = vec4(Lerp(result.rg, history.rg, hysteresis), 0.f, 1.f);
//////////////////////////////工程化修正/////////////////////////////////////////////
        imageStore(o_Texture, ivec3(gl_GlobalInvocationID), result);
        return;

    }else{
        // 边界
        memoryBarrier();        // 所有全局内存类型的屏障
        memoryBarrierShared();  // 对 shared 内存屏障
        barrier();   //这里有一个同步点 


        bool isCornerTexel = (LocalInvocationID.x == 0 || LocalInvocationID.x == (DDGI_PROBE_NUM_TEXELS_DISTANCE - 1)) && (LocalInvocationID.y == 0 || LocalInvocationID.y == (DDGI_PROBE_NUM_TEXELS_IRRANDIANCE - 1));
        bool isRowTexel = (LocalInvocationID.x > 0 && LocalInvocationID.x < (DDGI_PROBE_NUM_TEXELS_DISTANCE - 1));

        uvec3 copyCoordinates = uvec3(groupID.x * DDGI_PROBE_NUM_TEXELS_DISTANCE, groupID.y * DDGI_PROBE_NUM_TEXELS_DISTANCE, invocationID.z);

        if(isCornerTexel)
        {
            copyCoordinates.x += LocalInvocationID.x > 0 ? 1 : DDGI_PROBE_NUM_TEXELS_DISTANCE_INTERIOR;
            copyCoordinates.y += LocalInvocationID.y > 0 ? 1 : DDGI_PROBE_NUM_TEXELS_DISTANCE_INTERIOR;
        }
        else if(isRowTexel)
        {
            copyCoordinates.x += (DDGI_PROBE_NUM_TEXELS_DISTANCE - 1) - LocalInvocationID.x;
            copyCoordinates.y += LocalInvocationID.y + ((LocalInvocationID.y > 0) ? -1 : 1);
        }
        else // Column Texel
        {
            copyCoordinates.x += LocalInvocationID.x + ((LocalInvocationID.x > 0) ? -1 : 1);
            copyCoordinates.y += (DDGI_PROBE_NUM_TEXELS_DISTANCE - 1) - LocalInvocationID.y;
        }

        imageStore(o_Texture, ivec3(gl_GlobalInvocationID), imageLoad(o_Texture, ivec3(copyCoordinates)));
    }

}

#endif

IndirectLight

/*
    根据一个世界坐标，计算从探针中获取的Irrandiance
*/
vec3 DDGIGetIrrandianceByWorldPosition(vec3 worldPosition, vec3 direction, DDGISetting volume, texture2DArray IrrdianceTexture,texture2DArray distanceTexture){

    // 如果WorldPosition不在DDGIvolume内，直接返回
    // if(!DDGICheckPostion(worldPosition,volume)){
    //     return vec3(0.f, 0.f, 0.f);
    // }
    vec3  sumIrradiance = vec3(0.0f);
    float sumWeight = 0.0f;

    vec3 irradiance = vec3(0.f, 0.f, 0.f);
    float accumulatedWeights = 0.f;
    // 得到离worldPosition最近的探针坐标
    ivec3 baseProbeCoords = DDGIGetBaseProbeGridCoords(worldPosition, volume);
    vec3 baseProbeWorldPosition = DDGIGetProbeWorldPosition(baseProbeCoords,volume);
    vec3 alpha = clamp(((worldPosition - baseProbeWorldPosition) / volume.gridStep), vec3(0.f, 0.f, 0.f), vec3(1.f, 1.f, 1.f));

    for(int i = 0; i < 8; i++){
        float weight = 1.0;
        // 使用魔法获得一个探针😄
        ivec3  offset  = ivec3(i, i >> 1, i >> 2) & ivec3(1);
        ivec3  probeGridCoord   = clamp(baseProbeCoords + offset, ivec3(0), ivec3(volume.probeCount) - ivec3(1));
        uint probeIndex =  (probeGridCoord.y * DDGIGetProbesPerPlane(volume.probeCount)) + probeGridCoord.z * volume.probeCount.x + probeGridCoord.x;
        vec3 probePos = DDGIGetProbeWorldPosition(probeGridCoord, volume);

        // 方向系数
        {
            vec3 directionToProbe = normalize(probePos - worldPosition);
            weight *= Square(max(0.0001, (dot(directionToProbe, direction) + 1.0) * 0.5)) + 0.2;  
        }

        //切比雪夫系数
        {
            vec3 probeToPoint   = worldPosition - probePos;
            vec3 dir            = normalize(-probeToPoint);
            float dist          = length(probeToPoint);
            vec2 octantCoords = DDGIGetOctahedralCoordinates(dir);
            vec3 probeTextureUV = DDGIGetProbeUV(int(probeIndex), octantCoords, int(DDGI_PROBE_NUM_TEXELS_DISTANCE_INTERIOR), volume);
            vec2 temp = texture(sampler2DArray(distanceTexture,SAMPLER[0]),probeTextureUV).rg;  // 采样距离纹理
            float mean      = temp.x;
            float variance  = abs(Square(temp.x) - temp.y);

            float chebyshev = variance / (variance + Square(max(dist - mean, 0.0)));
            chebyshev       = max(Pow3(chebyshev), 0.0);  //以切比雪夫系数三次方作为权重

            weight *= (dist <= mean) ? 1.0 : chebyshev;
        }

        //避免计算精度问题
        {
            weight = max(0.000001, weight); 

            const float crushThreshold = 0.2f;
            if (weight < crushThreshold)
                weight *= weight * weight * (1.0f / Square(crushThreshold)); 
        }

        //三线性插值系数
        {
            vec3 trilinear = mix(1.0 - alpha, alpha, offset);
            weight *= trilinear.x * trilinear.y * trilinear.z;  
        }

        //采样，累计光照      
        {
            vec3 irradianceDir  = direction;
            vec2 octantCoords = DDGIGetOctahedralCoordinates(direction);
            vec3 probeTextureUV = DDGIGetProbeUV(int(probeIndex), octantCoords, int(DDGI_PROBE_NUM_TEXELS_IRRANDIANCE_INTERIOR), volume);
            vec3 probeIrradiance = texture(sampler2DArray(IrrdianceTexture,SAMPLER[0]),probeTextureUV).rgb;
            sumIrradiance += weight * probeIrradiance;
            sumWeight += weight;
        }

    }

    vec3 netIrradiance = sumIrradiance / sumWeight;
    return 2 * PI * netIrradiance; 
}

下图显示只渲染间接光

Math

首先回顾Irrandiance是单位半球上各个方向的Radiance的积分（×cos）

均匀分布的蒙特卡洛积分来近似上边这个积分式

DDGI存储一个Irrandiance时，并不是/N，而是/余弦权重之和，目的是减少方差

他的期望是N/2,对比蒙特卡洛积分还需要*2

总结

第一次实现比较复杂的全局光照效果，还是太逞能，总想抄RTXGI的各种实践Tick，实际根本理解不了。甚至还想着用Y-Up来构建Imager2DArray，导致逻辑十分混乱。大量时间用在了无用的调试Bug上。
还有一些DDGI的内容没有实现
1. RelocationVolumeProbes：读取一个探针交中背面的次数，如果大于阈值，说明Probe在墙里边，就朝着交中背面的方向偏移，或者交中正面的距离太短了，也进行偏移
2. ClassifyVolumeProbes：Classify主要是为了给probe进行分类，对于一些卡在墙中或者不在视野内的probe则不进行计算，以节省开销，当然，更重要的是，这个还能避免计算墙外probe，一定程度上减少漏光
总体来看第一次搞大动作比较失败~

游戏引擎开发实践（抗锯齿 TAA）

Thu, 04 Dec 2025 18:58:15 +0800

总体来看TAA思想比较简单易懂，实现思路更多是实践性的经验总结

TAA

TAA (Temporal Anti-Aliasing) 综合历史帧的数据来实现抗锯齿，这样会将每个像素点的多次采样均摊到多个帧中，相对的开销要小得多。在 MSAA 中，我们在一帧中，在每个像素中放置了多个次像素采样点。在 TAA 中，我们实现的方式，就是在每帧采样时，将采样的点进行偏移，实现抖动 (jitter)。

静态场景

在静态场景下，上一帧与本帧的颜色信息很好获得，只需要取相同位置进行采样，然后进行一个插值即可，然后用jitter进行抖动

1
2
3

float3 currColor = currBuffer.Load(pos);
float3 historyColor = historyBuffer.Load(pos);
return lerp(historyColor, currColor, 0.05f);

这里的取混合系数为0.05，意味着最新的一帧渲染结果，只对最终结果产生了5%的贡献。

将累计的过程展开来看的话，可以看出当前帧 TAA 后的结果，是包含了所有的历史帧结果的，说明这种方式是合理的：

在 TAA 中，我们实现的方式，就是在每帧采样时，将采样的点进行偏移，实现抖动 (jitter)。采样点抖动的偏移，是和 MSAA 的次像素采样点放置是相同的，都需要使用低差异的采样序列，来实现更好的抗锯齿效果。TAA 中都会直接使用 Halton 序列，采样的点位置如下所示：

1
2
3

// 先将Halton序列的值转化为 -0.5~0.5范围的偏移；再除以屏幕长度，得到UV下的偏移值；最后乘以2，是转化到裁剪空间中的偏移值
ProjectionMatrix.m02 += (HaltonSequence[Index].x - 0.5f) / ScreenWidth * 2;
ProjectionMatrix.m12 += (HaltonSequence[Index].y - 0.5f) / ScreenHeight * 2;

动态物体

要对历史数据进行混合，就要能够还原出当前物体在屏幕中投影的位置。为了能够精确地记录物体在屏幕空间中的移动，我们使用 Motion Vector 贴图来记录物体在屏幕空间中的变化距离，表示当前帧和上一帧中，物体在屏幕空间投影坐标的变化值

记录上一帧的投影矩阵和上一帧的模型位置，在本帧使用上一帧的透视变换逆变换就能把模型还原到上一帧的视角下，在上一帧视角下对比两帧模型的位置来拿到Motion Vector（这种有点麻烦，实战不用这种方案）

使用 MotionVector

已经知道模型的运动变化了，在本帧中复原模型到上一帧的位置，算出对应的屏幕坐标

// 减去抖动坐标值，得到当前实际的像素中心UV值
uv -= _Jitter;
// 减去Motion值，算出上帧的投影坐标
float2 uvLast = uv - motionVectorBuffer.Sample(point, uv);
//使用双线性模式采样
float3 historyColor = historyBuffer.Sample(linear, uvLast);

这样使用也是有问题的。比如一扇门，本来模型在门后，现在向右位移模型出现，但是按照MontionVector去采样上一帧的颜色时，其实采样到了门上。

因此为了得到更加平滑的数据，可以在当前像素点周围判断深度，取距离镜头最近的点位置，来采样 Motion Vector 的值，这样可以减弱遮挡错误的影响。

优化手法：使用 lerp 代替 if-else，可以在 GPU 并行计算中避免分支，提高效率

float2 GetClosestFragment(float2 uv)
{
    float2 k = _CameraDepthTexture_TexelSize.xy;
    //在上下左右四个点
    const float4 neighborhood = float4(
        SAMPLE_DEPTH_TEXTURE(_CameraDepthTexture, sampler_PointClamp, clamp(uv - k)),
        SAMPLE_DEPTH_TEXTURE(_CameraDepthTexture, sampler_PointClamp, clamp(uv + float2(k.x, -k.y))),
        SAMPLE_DEPTH_TEXTURE(_CameraDepthTexture, sampler_PointClamp, clamp(uv + float2(-k.x, k.y))),
        SAMPLE_DEPTH_TEXTURE(_CameraDepthTexture, sampler_PointClamp, clamp(uv + k))
    );
    // 获取离相机最近的点
    #if defined(UNITY_REVERSED_Z)
        #define COMPARE_DEPTH(a, b) step(b, a)
    #else
        #define COMPARE_DEPTH(a, b) step(a, b)
    #endif
    // 获取离相机最近的点，这里使用 lerp 是避免在shader中写分支判断
    float3 result = float3(0.0, 0.0, SAMPLE_DEPTH_TEXTURE(_CameraDepthTexture, sampler_PointClamp, uv));
    result = lerp(result, float3(-1.0, -1.0, neighborhood.x), COMPARE_DEPTH(neighborhood.x, result.z));
    result = lerp(result, float3( 1.0, -1.0, neighborhood.y), COMPARE_DEPTH(neighborhood.y, result.z));
    result = lerp(result, float3(-1.0,  1.0, neighborhood.z), COMPARE_DEPTH(neighborhood.z, result.z));
    result = lerp(result, float3( 1.0,  1.0, neighborhood.w), COMPARE_DEPTH(neighborhood.w, result.z));
    return (uv + result.xy * k);
}

//在周围像素中，寻找离相机最近的点
float2 closest = GetClosestFragment(uv);
//使用周围最近点，得到Velocity值，来计算上帧投影位置
float2 uvLast = uv - motionVectorBuffer.Sample(point, closest);
//...

另外对应历史点的采样，可以使用 Catmull–Rom 采样进行进一步优化，而不是使用默认的双线性采样

因为使用了双线性采样，所以得到的值会混合周围像素的颜色，造成结果略微模糊。如果想要使得到的历史结果质量更好，也可以使用一些特殊的过滤方式进行处理。比如UE4中使用 Catmull–Rom 的方式进行锐化过滤。Catmull-Rom方式的采样，会在目标点周围进行 5 次采样，然后根据相应权重进行过滤混合，额外的开销也非常大。

历史结果处理

由于像素抖动，模型变化，渲染光照变化导致渲染结果发生变化时，会导致历史帧得到的像素值失效，就会产生 鬼影/ghosting 和 闪烁 /flicking 问题。

为了缓解鬼影和闪烁的问题，我们还要对采样的历史帧和当前帧数据进行对比，将历史帧数据 clamp/截断 在合理的范围内。要确定当前帧目标像素的亮度范围，就需要读取当前帧数据目标像素周围 5 个或者 9 个像素点的颜色范围：

简单的做法就是直接进行 clamp：就是先在当前位置像素领域内判断颜色范围，之后采样历史点时把历史点的颜色clamp到这个范围内，防止它颜色变化太大

float3 AABBMin, AABBMax;
AABBMax = AABBMin = RGBToYCoCg(Color);
// 取得YCoCg色彩空间下，Clip的范围
for(int k = 0; k < 9; k++)
{
    float3 C = RGBToYCoCg(_MainTex.Sample(sampler_PointClamp, uv, kOffsets3x3[k]));
    AABBMin = min(AABBMin, C);
    AABBMax = max(AABBMax, C);
}
// 需要 Clip处理的历史数据
float3 HistoryYCoCg = RGBToYCoCg(HistoryColor);

// 简单地进行Clmap
float3 ResultYCoCg =  clmap(History, AABBMin, AABBMax);
//还原到RGB色彩空间，得到最终结果
HistoryColor.rgb = YCoCgToRGB(ResultYCoCg));

另外一种做法是进行 clip，clip的效果会更好，计算量也会相对较大二者的差别可从下图看出：

Clip和Clamp的区别，clamp简单粗暴地把颜色固定到范围角点上，clip则效果更好，首先获取范围内颜色的中间值，然后从历史点颜色发送射线到中间值，取射线与AABB盒交点位置的颜色值

混合结果

1
2
3

// 与上帧相比移动距离越远，就越倾向于使用当前的像素的值
blendFactor = saturate(0.05 + length(motion) * 100);
return lerp(historyColor, currColor, blendFactor);

因为我们用到了很多双线性采样，会使得得到的结果有些模糊，因此我们根据情况选择是否对结果进行一次简单的锐化。

实战

MotionVector

这块逻辑很清晰。因为模型在世界的位置只取决于模型实例的Model变换。而模型在屏幕像素的位置取决于投影矩阵和摄像机矩阵。所以每个模型都需要额外存储一个glm::mat4来存储上一帧的模型变换，摄像机也需要额外存储上一帧的投影矩阵（不过这个一般不会变），已经上一帧的View矩阵

在GBufferPass时，获取模型上一帧的Model矩阵，在FragShader中，根据像素存储插值出来的本帧Position和上帧Position进行计算，得到MotionVector

vec2 CalculateVelocity(vec4 pos, vec4 prevPos){
    vec4 curNDC  = CAMERAINFO.data.proj * CAMERAINFO.data.view * pos;
    curNDC /= curNDC.w;
    vec2 curUV = curNDC.xy * 0.5 + 0.5;
    vec4 prevNDC = CAMERAINFO.data.prevProj * CAMERAINFO.data.prevView * prevPos;
    prevNDC /= prevNDC.w;
    vec2 prevUV = prevNDC.xy * 0.5 + 0.5;
    return vec2(curUV.xy - prevUV.xy);  // MotionVector存储的是当前像素与上一帧中该位置信息所在的像素的偏移增量
}

混合历史信息

在TAApass，读取当前ViewPort，根据MotionVector把像素值偏移到上一帧位置，读取历史信息，混合历史信息和当前信息

#version 450 core
#include "../common/common.glsl"
#ifdef COMPUTE_SHADER

layout(set = 1,binding = 0,rgba32f) uniform image2D out_texture;
layout(set = 1,binding = 1) uniform texture2D velocityTexture;
layout(set = 1,binding = 2) uniform texture2D historyTexture;
layout(set = 1,binding = 3) uniform texture2D curTexture;
layout(set = 1,binding = 4) uniform texture2D depthTexture;


vec2 GetClosestFragment(vec2 uv)
{
    vec2 k = 1.0 / vec2(imageSize(out_texture)); 
    
    float depth0 = texture(sampler2D(depthTexture, SAMPLER[0]), clamp(uv - k, 0.0, 1.0)).r;
    float depth1 = texture(sampler2D(depthTexture, SAMPLER[0]), clamp(uv + vec2(k.x, -k.y), 0.0, 1.0)).r;
    float depth2 = texture(sampler2D(depthTexture, SAMPLER[0]), clamp(uv + vec2(-k.x, k.y), 0.0, 1.0)).r;
    float depth3 = texture(sampler2D(depthTexture, SAMPLER[0]), clamp(uv + k, 0.0, 1.0)).r;
    
    vec4 neighborhood = vec4(depth0, depth1, depth2, depth3);

    #define COMPARE_DEPTH(a, b) step(a, b)


    vec3 result = vec3(0.0, 0.0, texture(sampler2D(depthTexture, SAMPLER[0]), uv).r);
    
    result = mix(result, vec3(-1.0, -1.0, neighborhood.x), COMPARE_DEPTH(neighborhood.x, result.z));
    result = mix(result, vec3( 1.0, -1.0, neighborhood.y), COMPARE_DEPTH(neighborhood.y, result.z));
    result = mix(result, vec3(-1.0,  1.0, neighborhood.z), COMPARE_DEPTH(neighborhood.z, result.z));
    result = mix(result, vec3( 1.0,  1.0, neighborhood.w), COMPARE_DEPTH(neighborhood.w, result.z));

    #undef COMPARE_DEPTH
    return uv + result.xy * k;
}

layout(local_size_x = 16, local_size_y = 16, local_size_z = 1) in;
void main()
{
    // TODO: 此处的重投影逻辑可能在别的地方一会用到，应该写成单独的一个pass,把重投影后的UV写在一个纹理中
    ivec2 imgSize = ivec2(imageSize(out_texture));
    ivec2 invocID = ivec2(gl_GlobalInvocationID.xy);
    if (invocID.x >= imgSize.x || invocID.y >= imgSize.y) 
        return;
    vec2 texCoords = (vec2(invocID) + 0.5f) / vec2(imgSize);

    vec2 ClosestUv = GetClosestFragment(texCoords);

    vec2 velocity = texture(sampler2D(velocityTexture,SAMPLER[0]), ClosestUv).rg;  // 用采样点周围离摄像头最近的深度的UV来采样速度值
    vec2 historyTexCoords = texCoords - velocity;
    bool valid = historyTexCoords.x > 0.0f ? true : false;

    float blend = 0.95f;
    if(!valid)  blend = 1.0f;

    vec3 historyColor = texture(sampler2D(historyTexture,SAMPLER[0]), historyTexCoords).rgb;
    vec3 curColor = texture(sampler2D(curTexture,SAMPLER[0]), texCoords).rgb;
    vec3 mixColor = mix(curColor,historyColor, blend);
    imageStore(out_texture, invocID, vec4(mixColor, 1.0));
}

#endif

当然目前只是简单混合了颜色，渲染结果能看出来TAA的效果，但是移动物体或者摄像机时，鬼影十分严重

处理拖影

对比当前数据和历史数据，把历史数据clamp到合理范围内

#version 450 core
#include "../common/common.glsl"
#include "../common/TAA.glsl"
#include "../common/math.glsl"
#ifdef COMPUTE_SHADER

layout(set = 1,binding = 0,rgba32f) uniform image2D out_texture;
layout(set = 1,binding = 1) uniform texture2D velocityTexture;
layout(set = 1,binding = 2) uniform texture2D historyTexture;
layout(set = 1,binding = 3) uniform texture2D curTexture;
layout(set = 1,binding = 4) uniform texture2D depthTexture;

layout(local_size_x = 16, local_size_y = 16, local_size_z = 1) in;

void main()
{
    ivec2 invocID = ivec2(gl_GlobalInvocationID.xy);
    ivec2 imgSize = ivec2(imageSize(out_texture));
    if (invocID.x >= imgSize.x || invocID.y >= imgSize.y) 
        return;
    if(GetPostprocessSetting().TaaSetting.enable == 0){
        vec2 texCoords = (vec2(invocID) + 0.5f) / vec2(imgSize);
        vec3 curColor = texture(sampler2D(curTexture,SAMPLER[0]), texCoords).rgb;
        imageStore(out_texture, invocID, vec4(curColor, 1.0));
        return;
    }
    vec2 texelSize = 1.0 / vec2(imgSize);
    vec2 baseTexCoords = (vec2(invocID) + 0.5f) / vec2(imgSize);
    vec2 jitterUV = GetPostprocessSetting().TaaSetting.UVjetter * texelSize;
    vec2 texCoordsWithJitter = baseTexCoords + jitterUV;

    vec2 ClosestUv = GetClosestFragment(baseTexCoords, texelSize, depthTexture);
    vec2 velocity = texture(sampler2D(velocityTexture,SAMPLER[0]), ClosestUv).rg;
    vec2 historyTexCoords = baseTexCoords - velocity; 

    // 采样当前帧（带抖动）、历史帧（无抖动）颜色
    vec3 historyColor = texture(sampler2D(historyTexture,SAMPLER[0]), historyTexCoords).rgb;
    vec3 curColor = texture(sampler2D(curTexture,SAMPLER[0]), texCoordsWithJitter).rgb;

    const vec2 kOffsets3x3[9] = {
        vec2(-1.0, -1.0), vec2( 0.0, -1.0), vec2( 1.0, -1.0),
        vec2(-1.0,  0.0), vec2( 0.0,  0.0), vec2( 1.0,  0.0),
        vec2(-1.0,  1.0), vec2( 0.0,  1.0), vec2( 1.0,  1.0)
    };

    // fix 拖影：3x3邻域使用带抖动的UV，且钳位边界
    vec3 AABBMin = RGBToYCoCg(curColor);
    vec3 AABBMax = AABBMin;
    for(int k = 0; k < 9; k++)
    {
        vec2 uvOffset = kOffsets3x3[k] * texelSize;
        vec2 neighborUV = texCoordsWithJitter + uvOffset;
        neighborUV = clamp(neighborUV, 0.0, 1.0);
        vec3 C = RGBToYCoCg(texture(sampler2D(curTexture, SAMPLER[0]), neighborUV).rgb);
        AABBMin = min(AABBMin, C);
        AABBMax = max(AABBMax, C);
    }

    // Clip处理核心逻辑（无问题）
    vec3 HistoryYCoCg = RGBToYCoCg(historyColor);
    vec3 Filtered = (AABBMin + AABBMax) * 0.5f;
    vec3 RayOrigin = HistoryYCoCg;
    vec3 RayDir = Filtered - RayOrigin;
    float epsilon = 1.0 / 65536.0;
    RayDir = mix(vec3(epsilon), RayDir, greaterThan(abs(RayDir), vec3(epsilon)));
    vec3 InvRayDir = 1.0 / RayDir;
    vec3 MinIntersect = (AABBMin - RayOrigin) * InvRayDir;
    vec3 MaxIntersect = (AABBMax - RayOrigin) * InvRayDir;
    vec3 EnterIntersect = min(MinIntersect, MaxIntersect);
    float ClipBlend = max(EnterIntersect.x, max(EnterIntersect.y, EnterIntersect.z));
    ClipBlend = clamp(ClipBlend, 0.0, 1.0);
    vec3 ResultYCoCg = mix(HistoryYCoCg, Filtered, ClipBlend);
    historyColor = YCoCgToRGB(ResultYCoCg);

    float motionLength = length(velocity);
    float blendFactor = Saturate(0.05 + motionLength * 1000.0);
    bool valid = (historyTexCoords.x >= 0.0 && historyTexCoords.x <= 1.0) && 
                 (historyTexCoords.y >= 0.0 && historyTexCoords.y <= 1.0);
    if (!valid) {blendFactor = 1.0;}

    if (GetPostprocessSetting().TaaSetting.shaper == 1)
    {
        float strength = GetPostprocessSetting().TaaSetting.shaperStrength;
        vec2 topUV     = clamp(texCoordsWithJitter + vec2(0.0, -texelSize.y), 0.0, 1.0);
        vec2 leftUV    = clamp(texCoordsWithJitter + vec2(-texelSize.x, 0.0), 0.0, 1.0);
        vec2 rightUV   = clamp(texCoordsWithJitter + vec2(texelSize.x, 0.0), 0.0, 1.0);
        vec2 bottomUV  = clamp(texCoordsWithJitter + vec2(0.0, texelSize.y), 0.0, 1.0);

        vec3 topColor    = RGBToYCoCg(texture(sampler2D(curTexture, SAMPLER[0]), topUV).rgb);
        vec3 leftColor   = RGBToYCoCg(texture(sampler2D(curTexture, SAMPLER[0]), leftUV).rgb);
        vec3 centerColor = RGBToYCoCg(curColor);
        vec3 rightColor  = RGBToYCoCg(texture(sampler2D(curTexture, SAMPLER[0]), rightUV).rgb);
        vec3 bottomColor = RGBToYCoCg(texture(sampler2D(curTexture, SAMPLER[0]), bottomUV).rgb);

        vec3 sharpenSum = vec3(0.0);
        sharpenSum += (-1.0 * strength) * topColor;
        sharpenSum += (-1.0 * strength) * leftColor;
        sharpenSum += (5.0 * strength) * centerColor;
        sharpenSum += (-1.0 * strength) * rightColor;
        sharpenSum += (-1.0 * strength) * bottomColor;

        vec3 sharpenedColor = YCoCgToRGB(sharpenSum);
        sharpenedColor = max(sharpenedColor, vec3(0.0));
        curColor = sharpenedColor;
    }

    // HDR需要先转为LDRmix后再恢复， 如果不这样，目前测试发现会出现一个黑点，然后移动后越来越大
    historyColor = ToneMapping(max(historyColor, 0.0f));
    curColor = ToneMapping(max(curColor, 0.0f));
    vec3 mixColor = mix(historyColor, curColor, blendFactor);
    mixColor = InverseToneMapping(max(mixColor, 0.0f));

    imageStore(out_texture, invocID, vec4(mixColor, 1.0));
}

#endif

游戏引擎开发实践（高质量Bloom）

Mon, 10 Nov 2025 10:09:12 +0800

下采样

为什么要下采样

要想光晕扩的足够大，第一件事情就是扩大模糊的范围。一种非常简单的思路就是死命加大滤波盒的尺寸，使用一个巨大的 kernal 对纹理进行模糊。但是性能上肯定是吃不消，单 pass 的纹理采样次数是 N^2 而双 pass 是 N+N

此外还有一个问题，在处理高分辨率纹理时你需要等比地增加滤波盒的尺寸，才能形成同等大小的模糊。比如在 1000x1000 分辨率下用 250 像素的 kernal，模糊的结果占 1/4 屏幕，当分辨率增加到 2000x2000 的时候，要使用 500 像素的 kernal 才能达到同样的效果

回到模糊的问题，模糊滤波的本质是查询 kernal 范围内的所有像素并加权平均，即范围查询问题。在计算机图形学中实现快速范围查询，通常会请到老朋友 Mipmap 出场。Mipmap 将图像大小依次折半形成金字塔，mip[i] 中的单个像素代表了 mip[i-1] 中的 2x2 像素块均值，也代表 mip[i-2] 中的 4x4 像素块均值：

为什么不直接平均，而是用高斯模糊

为什么不直接用平均，而是用高斯模糊。下图为直接使用mipmap，进行下采样

下图为高斯模糊，明显更加圆滑，直接MipMap还是方形的

所以进行下采样时，使用高斯模糊，而不是普通的平均。这只是第一步

直接将最高层级 mip 叠加到图像上虽然能够产生足够大的光晕扩散，但是发光物的中心区域不够明亮。此外，发光物和泛光之间没有过度而是直接跳变

亮度跳变导致的

为了实现发光物和最高层 mip 之间的过渡，我们需要叠加所有的 mip 层级到原图上。因为 mip[i] 是基于 mip[i-1] 进行计算的，相邻层级之间相对连续则不会产生跳变

上采样

引擎实践

游戏引擎开发实践（大气渲染）

Fri, 07 Nov 2025 18:08:43 +0800

其实OpenGL渲染器已经完成的单次散射的大气渲染，但是对原理理解不透彻，这次准备更完善一点，并且实现整套的多重散射大气渲染，参考UE的实现，以及几篇论文

Precomputed Atmospheric Scattering

Eric Bruneton, Fabrice Neyret在2008年发表的论文，在2017年有了一个示例代码

Transmittance 透射率

基本概念：透射是可乘的

对于 三点共线 $p, q, r$（顺序为 $p \to q \to r$）
从 p 到 r 的透射率 = 从 p 到 q 的透射 × 从 q 到 r 的透射

$$ T(p,r) = T(p,q) \cdot T(q,r) $$

这是光线传输的乘法性质，因为大气衰减是指数衰减，连续段可以相乘。

处理大气边界

对于点对 $p, q$
将从 $p$ 出发到 $q$ 的射线延长，找到它 与大气上下边界的最近交点 $i$
透射可以表示为：

$$ T(p,q) = \frac{T(p,i)}{T(q,i)} $$

特殊情况：如果 $[p,q]$ 与地面相交 → 透射为 0（光被阻挡）
直观理解：你只需要知道从大气内部某点到大气边界的透射，再通过除法就能得到任意点对的透射。

透射率的预计算

离地心的距离 r 和天顶角的余弦$\mu = \cos\theta$这两个参数，就可以描述从任意点出发沿着任意方向到大气层的路径上的传输衰减。

所以预计算的是摄像机在任意高度，不同抬头角度看向大气层的边界的投射率

// TODO: 知乎https://zhuanlan.zhihu.com/p/595576594上对坐标到纹理的重映射方法与论文不一样

单次散射

光路即使没有经过太阳，太阳也对这条光路有贡献，因为本来不应该经过视线的光由于空气粒子的反弹进入了视线，形成光路

光线在空气中主要会发生两种物理现象：散射和透射。

散射：散射是指一束光和大气层中的微粒发生碰撞之后，众多粒子各奔东西四散而逃的物理现象。我们把逃逸到视线方向进而产生颜色贡献的光照能量记作：S
透射（Transmittance）描述的是光在介质中穿行所造成的能量衰减，记作：$I = T(I_0)$

所以单次散射的流程：

T1透射（衰减）
S点散射
T2透射（衰减）

事实上在大气层内太阳光可以被看做是平行光，因此需要对视线方向做积分来重建无数条光路的总和：

散射理论

解决散射系数和相位函数是什么

在大气渲染中也是类似的，我们通过 “散射系数” 和 “相位函数” 来描述光线散射的现象。散射系数$\sigma$ 描述了光线和空气分子发生一次反弹之后逃逸到各个方向的（失去的）能量总和

剩下的能量并非完全进入我们的视线，而是会四下逃窜。因此相位函数 $phase(\theta)$ 表述了反弹后剩下的能量有多少能够逃到指定的方向上。相位函数和 BRDF 一样在定义域上积分等于 1.0，这保证了能量守恒：

理解就是：1. 散射系数决定有多少能量会进行散射 2. 相位函数决定这些散射的能量有多少会沿着$\theta$方向进行散射（注意散射的能量才是渲染用的，而不是剩下的）

将散射系数和相位函数简单相乘就能得到在某点发生一次散射之后，逃逸到某个方向上的能量大小

$S = \sigma \cdot \text{phase}(\theta)$

大气层并非完全均匀，大气分子的密度随着高度的增加而减少。越少的分子数目意味着发生散射的概率越低。此外大气对红、黄、蓝三种波长的光有着不同的散射量。因此散射系数通常是波长和高度的函数

$\sigma(\lambda, h)$

对于空气这种介质，它的密度通常随着海拔的增高而降低。密度的高低反应在数值上就是散射概率的减少，散射概率的减少对应着散射后剩余能量的增加。因此可以用海平面（Y=0）处的散射系数和海平面 h 处的高度密度衰减函数来描述任意高度的散射系数

$\sigma(\lambda, h) = \sigma(\lambda, 0)\rho(h)$

单次散射预计算

这篇论文使用的是4D表，属于是把计算结果全部打表了

单次散射函数依赖 4 个参数：

$$ (r, \mu, \mu_s, \nu) $$

$r$：观察点到地心的距离
$\mu = \cos(\theta_v)$：视线方向与地心方向余弦
$\mu_s = \cos(\theta_s)$：太阳方向与地心方向余弦
$\nu = \cos(\theta_{vs})$：视线方向与太阳方向余弦

多重散射

二次散射的计算需要一次散射，三次需要二次，所以一层一层往上算来得到多重散射效果。

太阳光可能散射两次才散射到ViewDir

沿着ViewDir上某一点的计算不仅要计算来自太阳的散射，而是一个球面积分（也就是说任意方向都可能散射到这个点），对于球面上任何一点的贡献计算又是一层积分运算（球面积分的每一个采样方向，都要逐步进行 raymarch 到大气层边缘。如下图最右边蓝色线条）。所有2次多重散射就需要3重积分来求解

设：

$L^{(1)}(p, \omega)$：点 p 向方向 ω 的单次散射光。
$L^{(2)}(p, \omega)$：点 p 向方向 ω 的二次散射光。
$L^{(3)}(p, \omega)$：点 p 向方向 ω 的三次散射光。

那么：

$$ L^{(2)}(p, \omega) = \int_{q, \omega'} L^{(1)}(q, \omega') \cdot f_s(\omega', \omega) $$

以此类推：

$$ L^{(n)}(p, \omega) = \int_{q, \omega'} L^{(n-1)}(q, \omega') \cdot f_s(\omega', \omega) $$

所以：

计算第 n 阶散射，必须先有第 n−1 阶的结果。

有点懂，但没完全懂….

A Scalable and Production Ready Sky and Atmosphere Rendering Technique

2020年Epic发表介绍大气渲染的论文

多重散射的优化

这篇论文主要对多重散射需要一层一层计算迭代来实现的情况做了简化。作者通过做出一些对结果影响不大的简化假设，大幅度地降低了Multiple Scattering计算的消耗，并且使用数值方法计算无穷阶Scattering之和变为可能。

论文提到的Sky-View LUT就是最终渲染结果存储的纹理

TODO: 在靠近地平线部分大气散射会变得高频，因此UV的映射不是线性，而是靠近地平线部分会分配更多像素

大气透视查找表

除了天空，远处被大气覆盖的物体也要受到大气的影响。远处的物体对相机的贡献主要来自于两点，物体自身的颜色乘以着色点到相机的 transmittance，以及相机到着色点这段路径中间受到的大气散射。对应下图两条不同颜色的路径：

分割Camera Frustum（和Cluster Rendering中的分割一样），计算每个格子到Camera方向的In-Scattering和Transmittance，保存在Volume Texture中；
在Volume Texture的z轴方向上根据Transmittance累加In-Scattering，使得每一个单元格保存的是该单元格到Camera的Luminance；
在Opaque渲染之后，做一次Post Processing，采样上述的Volume Texture，对场景中的物体添加Aerial Perspective；
Transparent物体渲染时在VS中采样Volume Texture添加Aerial Perspective。

引擎实践

Transmittance Lut

用ComputerShader预计算一张2D表，记录每个高度，每个天顶角到大气外圈的透射率。

流程是把UV坐标映射到X：不同天顶角余弦值（-1，1）,Y：不同高度（0，1）用得到的数据计算透射率后保存

#version 450 core
#ifdef COMPUTE_SHADER
#define LOCAL_SIZE 8
#include "include/SkyCommon.glslh"

layout(rgba32f, binding = 0) uniform writeonly image2D transmittanceLut;


layout(local_size_x = LOCAL_SIZE, local_size_y = LOCAL_SIZE, local_size_z = 1) in;
void main()
{
    ivec2 texelCoord = ivec2(gl_GlobalInvocationID.xy);

	AtmosphereParameter Atmosphere;  // TODO: AtmosphereUniform
    Atmosphere.PlanetRadius = 6360000.0;
    Atmosphere.AtmosphereHeight = 60000.0;
    Atmosphere.RayleighScatteringScalarHeight = 8000.0;
    Atmosphere.MieScatteringScalarHeight = 1200.0;
    Atmosphere.OzoneLevelCenterHeight = 25000.0;
    Atmosphere.OzoneLevelWidth = 15000.0;


	float bottomRadius = Atmosphere.PlanetRadius;
	float topRadius = Atmosphere.PlanetRadius + Atmosphere.AtmosphereHeight;


    ivec2 lutSize = imageSize(transmittanceLut);
    float viewHeight;
    float viewZenithCosAngle;
    vec2 uv = vec2(texelCoord) / vec2(lutSize);
	float r = 0.0;

    UvToTransmittanceLutParams(bottomRadius, topRadius, uv, viewZenithCosAngle, r);
	float sin_theta = sqrt(1.0 - viewZenithCosAngle * viewZenithCosAngle);
	vec3 CameraPosition = vec3(0.0, r, 0.0);
	vec3 viewDir = vec3(sin_theta, viewZenithCosAngle, 0);

	float dis = RayIntersectSphere(vec3(0,0,0), topRadius, CameraPosition, viewDir);
	vec3 hitPoint = CameraPosition + viewDir * dis;
	vec3 color = Transmittance(Atmosphere, CameraPosition, hitPoint);
    imageStore(transmittanceLut, texelCoord, vec4(color, 1.0));
}

#endif

// UV转LUT参数（高度、天顶角的余弦值）
void UvToTransmittanceLutParams(float bottomRadius, float topRadius, vec2 uv, out float mu, out float r)
{
    float x_mu = uv.x;
    float x_r = uv.y;

    float H = sqrt(max(0.0f, topRadius * topRadius - bottomRadius * bottomRadius));
    float rho = H * x_r;
    r = sqrt(max(0.0f, rho * rho + bottomRadius * bottomRadius));

    float d_min = topRadius - r;
    float d_max = rho + H;
    float d = d_min + x_mu * (d_max - d_min);
    mu = d == 0.0f ? 1.0f : (H * H - rho * rho - d * d) / (2.0f * r * d);
    mu = clamp(mu, -1.0f, 1.0f);
}

最终生成的lut如下，在Vulkan中，0,0在右上角，所以和参考博客生成的lutY轴反转了

SkyView Lut

用一张2D（球面坐标）纹理存储不同ViewDir下的大气渲染结果

#version 450 core
#ifdef COMPUTE_SHADER
#define LOCAL_SIZE 8
#include "include/SkyCommon.glslh"
// 二维 UV 坐标当作球面坐标 (θ, φ) 映射到单位球方向向量。 一个技巧：用2D球面来存储3D的方向向量了
vec3 UVToViewDir(vec2 uv)
{
    float theta = (1.0 - uv.y) * PI;
    float phi = (uv.x * 2 - 1) * PI;
    
    float x = sin(theta) * cos(phi);
    float z = sin(theta) * sin(phi);
    float y = cos(theta);

    return vec3(x, y, z);
}
layout(rgba32f, binding = 0) uniform writeonly image2D SkyViewLut;
layout(binding = 1) uniform sampler2D u_TransmittanceLut;
layout(binding = 2) uniform sampler2D u_MultiScatteringLut;
layout(std140, set = 0, binding = 3) uniform SceneData
{
	DirectionalLight DirectionalLights;
	float EnvironmentMapIntensity;
} u_Scene;
layout(set = 0,binding = 4) uniform CameraDataUniform {
    mat4 view;
    mat4 proj;
	mat4 viewProj;
	float width;
	float height;
	float Near;
	float Far;
	vec3 CameraPosition;
} u_CameraData;
layout(local_size_x = LOCAL_SIZE, local_size_y = LOCAL_SIZE, local_size_z = 1) in;
void main()
{
    ivec2 texelCoord = ivec2(gl_GlobalInvocationID.xy);

	AtmosphereParameter Atmosphere = BuildAtmosphereParameter();

	ivec2 lutSize = imageSize(SkyViewLut);

	vec2 uv = vec2(texelCoord) / vec2(lutSize);
    vec3 viewDir = UVToViewDir(uv);
	vec3 lightDir = normalize(-u_Scene.DirectionalLights.Direction);
	vec3 cameraPosition = normalize(u_CameraData.CameraPosition);

	float h = cameraPosition.y - Atmosphere.SeaLevel + Atmosphere.PlanetRadius;
	vec3 eyePos = vec3(0, h, 0);

	vec3 color = GetSkyView(Atmosphere, eyePos, viewDir, lightDir, -1.0f,u_TransmittanceLut, u_MultiScatteringLut);

    imageStore(SkyViewLut, texelCoord, vec4(color, 1.0));

}

#endif

vec3 GetSkyView(
    in AtmosphereParameter param, vec3 eyePos, vec3 viewDir, vec3 lightDir, float maxDis, 
    sampler2D _transmittanceLut, sampler2D _multiScatteringLut)
{
    const int N_SAMPLE = 32;
    vec3 color = vec3(0, 0, 0);

    // 光线和大气层, 星球求交
    float dis = RayIntersectSphere(vec3(0,0,0), param.PlanetRadius + param.AtmosphereHeight, eyePos, viewDir);
    float d = RayIntersectSphere(vec3(0,0,0), param.PlanetRadius, eyePos, viewDir);
    if(dis < 0) return color; 
    if(d > 0) dis = min(dis, d);
    if(maxDis > 0) dis = min(dis, maxDis);  // 带最长距离 maxDis 限制, 方便 aerial perspective lut 部分复用代码

    float ds = dis / float(N_SAMPLE);
    vec3 p = eyePos + (viewDir * ds) * 0.5;
    vec3 sunLuminance = param.SunLightColor * param.SunLightIntensity;
    vec3 opticalDepth = vec3(0, 0, 0);

    for(int i=0; i<N_SAMPLE; i++)
    {
        // 积累沿途的湮灭系数
        float h = length(p) - param.PlanetRadius;
        vec3 extinction = RayleighCoefficient(param, h) + MieCoefficient(param, h) +  // scattering
                            OzoneAbsorption(param, h) + MieAbsorption(param, h);        // absorption
        opticalDepth += extinction * ds;

        vec3 t1 = TransmittanceToAtmosphere(param, p, lightDir, _transmittanceLut);
        vec3 s  = Scattering(param, p, lightDir, viewDir);
        vec3 t2 = exp(-opticalDepth);
        
        // 单次散射
        vec3 inScattering = t1 * s * t2 * ds * sunLuminance;
        color += inScattering;

        // 多重散射
        vec3 multiScattering = GetMultiScattering(param, p, lightDir, _multiScatteringLut);
        color += multiScattering * t2 * ds * sunLuminance;

        p += viewDir * ds;
    }

    return color;
}

游戏引擎开发实践（Vulkan架构篇：Image）

Sun, 02 Nov 2025 20:46:56 +0800

这里涉及的整套结构已经废弃，转为研究更轻薄的RHI层，放弃过度封装！

本文主要讲解GameEngine项目Vulkan封装，在渲染层只需要调用RHI层接口即可

总览

所有底层对象创建都需要使用xxxSpecification来说明创建需求，调用RHI接口函数，创建底层Vulkan对象

Image

enum class ImageFormat
{
    None = 0,
    RED8UN,
    RED8UI,
    ...
    // Defaults
    Depth = DEPTH32F,
};
// 图片用途
enum class ImageUsage
{
	None = 0,
	Texture,
	Attachment,
	Storage,
	HostRead
};
// 图片说明书
struct ImageSpecification
{
	std::string DebugName;
	ImageFormat Format = ImageFormat::RGBA;  // 图片格式
	ImageUsage Usage = ImageUsage::Texture; // 用途
	bool Transfer = false; // 是否需要转移
	uint32_t Width = 1;
	uint32_t Height = 1;
	uint32_t Mips = 1; // mip多少层，在Vulkan中也需要自己指定
	uint32_t Layers = 1; // Layer，图层通常用于立方体纹理（6 个面）、数组纹理（多个图像的集合）等场景。
	bool CreateSampler = true;
};

class Image : public RendererResource
{
public:
	virtual ~Image() = default;

	virtual void Resize(const uint32_t width, const uint32_t height) = 0;
	virtual void Invalidate() = 0;
	virtual void Release() = 0;

	virtual uint32_t GetWidth() const = 0;
	virtual uint32_t GetHeight() const = 0;
	virtual glm::uvec2 GetSize() const = 0;
	virtual bool HasMips() const = 0;

	virtual float GetAspectRatio() const = 0;

	virtual ImageSpecification& GetSpecification() = 0;
	virtual const ImageSpecification& GetSpecification() const = 0;

	virtual Buffer GetBuffer() const = 0;
	virtual Buffer& GetBuffer() = 0;

	virtual uint64_t GetGPUMemoryUsage() const = 0; 

	virtual void CreatePerLayerImageViews() = 0;

	virtual uint64_t GetHash() const = 0;

	virtual void SetData(Buffer buffer) = 0;
	virtual void CopyToHostBuffer(Buffer& buffer) const = 0; 
};
class Image2D : public Image
{
public:
    static Ref<Image2D> Create(const ImageSpecification& specification, Buffer buffer = Buffer());
    virtual void Resize(const glm::uvec2& size) = 0;
    virtual bool IsValid() const = 0;
};

到VulkanImage2D后，创建过程

处理图片用途

声明图像的所有用途（需用按位或组合），类型为 VkImageUsageFlags，常见取值：
- VK_IMAGE_USAGE_SAMPLED_BIT：可被采样器访问（作为纹理）。
- VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT：作为颜色附着（渲染目标）。
- VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT：作为深度 / 模板附着。
- VK_IMAGE_USAGE_TRANSFER_SRC_BIT/VK_IMAGE_USAGE_TRANSFER_DST_BIT：作为复制操作的源 / 目标。
- VK_IMAGE_USAGE_STORAGE_BIT：作为存储图像（Compute Shader 读写）。
用途需与后续操作匹配，否则会触发验证层错误。

VkImageUsageFlags usage = VK_IMAGE_USAGE_SAMPLED_BIT;
if (m_Specification.Usage == ImageUsage::Attachment) 
{
	if (Utils::IsDepthFormat(m_Specification.Format))
		usage |= VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT;
	else
		usage |= VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
}
if (m_Specification.Transfer || m_Specification.Usage == ImageUsage::Texture)
{
	usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT;
}
if (m_Specification.Usage == ImageUsage::Storage)
{
	usage |= VK_IMAGE_USAGE_STORAGE_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT;
}

VkImageCreateInfo

VkImageCreateInfo imageCreateInfo = {};
imageCreateInfo.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
imageCreateInfo.imageType = VK_IMAGE_TYPE_2D;
imageCreateInfo.format = vulkanFormat;
imageCreateInfo.extent.width = m_Specification.Width;
imageCreateInfo.extent.height = m_Specification.Height;
imageCreateInfo.extent.depth = 1;
imageCreateInfo.mipLevels = m_Specification.Mips;
imageCreateInfo.arrayLayers = m_Specification.Layers;
imageCreateInfo.samples = VK_SAMPLE_COUNT_1_BIT;
imageCreateInfo.tiling = m_Specification.Usage == ImageUsage::HostRead ? VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL; // 内存布局
imageCreateInfo.usage = usage;
m_Info.MemoryAlloc = allocator.AllocateImage(imageCreateInfo, memoryUsage, m_Info.Image, &m_GPUAllocationSize);
s_ImageReferences[m_Info.Image] = this;
VKUtils::SetDebugUtilsObjectName(device, VK_OBJECT_TYPE_IMAGE, m_Specification.DebugName, m_Info.Image);

VkImageViewCreateInfo（默认创建的ImageView是包含Image所有（所有mip，所有layer）的ImageView）

VkImageViewCreateInfo imageViewCreateInfo = {};
imageViewCreateInfo.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO;
imageViewCreateInfo.viewType = m_Specification.Layers > 1 ? VK_IMAGE_VIEW_TYPE_2D_ARRAY : VK_IMAGE_VIEW_TYPE_2D;
imageViewCreateInfo.format = vulkanFormat;
imageViewCreateInfo.flags = 0;
imageViewCreateInfo.subresourceRange = {};
imageViewCreateInfo.subresourceRange.aspectMask = aspectMask;
imageViewCreateInfo.subresourceRange.baseMipLevel = 0;
imageViewCreateInfo.subresourceRange.levelCount = m_Specification.Mips;
imageViewCreateInfo.subresourceRange.baseArrayLayer = 0;
imageViewCreateInfo.subresourceRange.layerCount = m_Specification.Layers;
imageViewCreateInfo.image = m_Info.Image;
VK_CHECK_RESULT(vkCreateImageView(device, &imageViewCreateInfo, nullptr, &m_Info.ImageView));
VKUtils::SetDebugUtilsObjectName(device, VK_OBJECT_TYPE_IMAGE_VIEW, fmt::format("{} default image view", m_Specification.DebugName), m_Info.ImageView);

采样器

if (m_Specification.CreateSampler)
{
	VkSamplerCreateInfo samplerCreateInfo = {};
	samplerCreateInfo.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO;
	samplerCreateInfo.maxAnisotropy = 1.0f;
	if (Utils::IsIntegerBased(m_Specification.Format))
	{
		samplerCreateInfo.magFilter = VK_FILTER_NEAREST;
		samplerCreateInfo.minFilter = VK_FILTER_NEAREST;
		samplerCreateInfo.mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST;
	}
	else
	{
		samplerCreateInfo.magFilter = VK_FILTER_LINEAR;
		samplerCreateInfo.minFilter = VK_FILTER_LINEAR;
		samplerCreateInfo.mipmapMode = VK_SAMPLER_MIPMAP_MODE_LINEAR;
	}

	samplerCreateInfo.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE;
	samplerCreateInfo.addressModeV = samplerCreateInfo.addressModeU;
	samplerCreateInfo.addressModeW = samplerCreateInfo.addressModeU;
	samplerCreateInfo.mipLodBias = 0.0f;
	samplerCreateInfo.minLod = 0.0f;
	samplerCreateInfo.maxLod = 100.0f;
	samplerCreateInfo.borderColor = VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE;
	m_Info.Sampler = Vulkan::CreateSampler(samplerCreateInfo);
	VKUtils::SetDebugUtilsObjectName(device, VK_OBJECT_TYPE_SAMPLER, fmt::format("{} default sampler", m_Specification.DebugName), m_Info.Sampler);
}

设置图片默认布局 Storage和HostRead才会设置，其他图片都还是无布局状态，不过RenderPass设置可以设置附件的开始和结束布局

if (m_Specification.Usage == ImageUsage::Storage) // General
{
	// Transition image to GENERAL layout
	VkCommandBuffer commandBuffer = VulkanContext::GetCurrentDevice()->GetCommandBuffer(true);

	VkImageSubresourceRange subresourceRange = {};
	subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
	subresourceRange.baseMipLevel = 0;
	subresourceRange.levelCount = m_Specification.Mips;
	subresourceRange.layerCount = m_Specification.Layers;

	Utils::InsertImageMemoryBarrier(commandBuffer, m_Info.Image,
		0, 0,
		VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL,
		VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
		subresourceRange);

	VulkanContext::GetCurrentDevice()->FlushCommandBuffer(commandBuffer);
}
else if (m_Specification.Usage == ImageUsage::HostRead) // VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
{
	// Transition image to TRANSFER_DST layout
	VkCommandBuffer commandBuffer = VulkanContext::GetCurrentDevice()->GetCommandBuffer(true);

	VkImageSubresourceRange subresourceRange = {};
	subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
	subresourceRange.baseMipLevel = 0;
	subresourceRange.levelCount = m_Specification.Mips;
	subresourceRange.layerCount = m_Specification.Layers;

	Utils::InsertImageMemoryBarrier(commandBuffer, m_Info.Image,
		0, 0,
		VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
		VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
		subresourceRange);

	VulkanContext::GetCurrentDevice()->FlushCommandBuffer(commandBuffer); // 会把命令提交并等待完成
}

资源描述符信息封装（用默认的构造写一个资源描述符）

	void VulkanImage2D::UpdateDescriptor()
	{
		if (m_Specification.Format == ImageFormat::DEPTH24STENCIL8 ||
			m_Specification.Format == ImageFormat::DEPTH32F ||
			m_Specification.Format == ImageFormat::DEPTH32FSTENCIL8UINT) {
			m_DescriptorImageInfo.imageLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_READ_ONLY_OPTIMAL;
		}
		else if (m_Specification.Usage == ImageUsage::Storage) {
			m_DescriptorImageInfo.imageLayout = VK_IMAGE_LAYOUT_GENERAL;
		}
		else if (m_Specification.Usage == ImageUsage::HostRead) {
			m_DescriptorImageInfo.imageLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;
		}
		else {
			m_DescriptorImageInfo.imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
		}

		m_DescriptorImageInfo.imageView = m_Info.ImageView;

		m_DescriptorImageInfo.sampler = m_Info.Sampler;
	}

另外Image类需要提供创建不同Mip不同Layer的ImageView的功能，用于不同的场景，比如Mip用于使用ComputerShader创建HZB，Layer用于CSM级联阴影，用一个Image的不同Layer表示不同的级联

Texture

Texture就是把Image再封装一层，用于纹理采样，另外封装了2D纹理和Cube纹理，后续还可以封装3D纹理，用于体积云等等

	struct TextureSpecification
	{
		ImageFormat Format = ImageFormat::RGBA;
		uint32_t Width = 1;
		uint32_t Height = 1;
		TextureWrap SamplerWrap = TextureWrap::Repeat;
		TextureFilter SamplerFilter = TextureFilter::Linear;

		bool GenerateMips = true;
		bool Storage = false;
		bool StoreLocally = false;

		std::string DebugName;
	};
	class Texture : public RendererResource
	{
	public:
		virtual ~Texture() {}

		virtual void Bind(uint32_t slot = 0) const = 0;

		virtual ImageFormat GetFormat() const = 0;
		virtual uint32_t GetWidth() const = 0;
		virtual uint32_t GetHeight() const = 0;
		virtual glm::uvec2 GetSize() const = 0;

		virtual uint32_t GetMipLevelCount() const = 0;
		virtual std::pair<uint32_t, uint32_t> GetMipSize(uint32_t mip) const = 0;

		virtual uint64_t GetHash() const = 0;

		virtual TextureType GetType() const = 0;
	};

	class Texture2D : public Texture
	{
	public:
		static Ref<Texture2D> Create(const TextureSpecification& specification);
		static Ref<Texture2D> Create(const TextureSpecification& specification, const std::filesystem::path& filepath);
		static Ref<Texture2D> Create(const TextureSpecification& specification, Buffer imageData);
		virtual void CreateFromFile(const TextureSpecification& specification, const std::filesystem::path& filepath) = 0;
		virtual void CreateFromBuffer(const TextureSpecification& specification, Buffer data = Buffer()) = 0;
		virtual void ReplaceFromFile(const TextureSpecification& specification, const std::filesystem::path& filepath) = 0;

		virtual void Resize(const glm::uvec2& size) = 0;
		virtual void Resize(const uint32_t width, const uint32_t height) = 0;

		virtual Ref<Image2D> GetImage() const = 0;

		virtual void Lock() = 0;
		virtual void Unlock() = 0;

		virtual Buffer GetWriteableBuffer() = 0;

		virtual bool Loaded() const = 0;

		virtual const std::filesystem::path& GetPath() const = 0;

		virtual TextureType GetType() const override { return TextureType::Texture2D; }

		static AssetType GetStaticType() { return AssetType::Texture; }
		virtual AssetType GetAssetType() const override { return GetStaticType(); }
	};
	class TextureCube : public Texture
	{
	public:
		static Ref<TextureCube> Create(const TextureSpecification& specification, Buffer imageData = Buffer());

		virtual TextureType GetType() const override { return TextureType::TextureCube; }
		virtual void GenerateMips(bool readonly = false) = 0;
		virtual void GenerateMips(Ref<RenderCommandBuffer> renderCommandBuffer,bool readonly = false) = 0;
		static AssetType GetStaticType() { return AssetType::EnvMap; }
		virtual AssetType GetAssetType() const override { return GetStaticType(); }
	};

创建底层Image对象

ImageSpecification imageSpec;
imageSpec.Format = m_Specification.Format;
imageSpec.Width = m_Specification.Width;
imageSpec.Height = m_Specification.Height;
imageSpec.Mips = specification.GenerateMips ? GetMipLevelCount() : 1;
imageSpec.DebugName = specification.DebugName;
imageSpec.CreateSampler = false;
m_Image = Image2D::Create(imageSpec);

从文件中读取数据

`1`	`m_ImageData = TextureImporter::ToBufferFromFile(filepath, m_Specification.Format, m_Specification.Width, m_Specification.Height);`

数据写入Image对象（先把数据放在临时缓冲区，再转移到图片），生成MipMap后，转为VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL布局

void VulkanTexture2D::SetData(Buffer buffer)
{
	auto device = VulkanContext::GetCurrentDevice();
	Ref<VulkanImage2D> image = m_Image.As<VulkanImage2D>();
	auto& info = image->GetImageInfo();

	VkDeviceSize size = m_ImageData.Size;

	VkMemoryAllocateInfo memAllocInfo{};
	memAllocInfo.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO;

	VulkanAllocator allocator("Texture2D");

	// Create staging buffer
	VkBufferCreateInfo bufferCreateInfo{};
	bufferCreateInfo.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;
	bufferCreateInfo.size = size;
	bufferCreateInfo.usage = VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
	bufferCreateInfo.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
	VkBuffer stagingBuffer;
	VmaAllocation stagingBufferAllocation = allocator.AllocateBuffer(bufferCreateInfo, VMA_MEMORY_USAGE_CPU_TO_GPU, stagingBuffer);

	// Copy data to staging buffer
	uint8_t* destData = allocator.MapMemory<uint8_t>(stagingBufferAllocation);
	HZ_CORE_ASSERT(m_ImageData.Data);
	memcpy(destData, m_ImageData.Data, size);
	allocator.UnmapMemory(stagingBufferAllocation);

	VkCommandBuffer copyCmd = device->GetCommandBuffer(true);

	// Image memory barriers for the texture image

	// The sub resource range describes the regions of the image that will be transitioned using the memory barriers below
	VkImageSubresourceRange subresourceRange = {};
	// Image only contains color data
	subresourceRange.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
	// Start at first mip level
	subresourceRange.baseMipLevel = 0;
	subresourceRange.levelCount = 1;
	subresourceRange.layerCount = 1;

	// Transition the texture image layout to transfer target, so we can safely copy our buffer data to it.
	VkImageMemoryBarrier imageMemoryBarrier{};
	imageMemoryBarrier.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER;
	imageMemoryBarrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
	imageMemoryBarrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
	imageMemoryBarrier.image = info.Image;
	imageMemoryBarrier.subresourceRange = subresourceRange;
	imageMemoryBarrier.srcAccessMask = 0;
	imageMemoryBarrier.dstAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT;
	imageMemoryBarrier.oldLayout = VK_IMAGE_LAYOUT_UNDEFINED;
	imageMemoryBarrier.newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL;

	// Insert a memory dependency at the proper pipeline stages that will execute the image layout transition
	// Source pipeline stage is host write/read execution (VK_PIPELINE_STAGE_HOST_BIT)
	// Destination pipeline stage is copy command execution (VK_PIPELINE_STAGE_TRANSFER_BIT)
	vkCmdPipelineBarrier(
		copyCmd,
		VK_PIPELINE_STAGE_HOST_BIT,
		VK_PIPELINE_STAGE_TRANSFER_BIT,
		0,
		0, nullptr,
		0, nullptr,
		1, &imageMemoryBarrier);

	VkBufferImageCopy bufferCopyRegion = {};
	bufferCopyRegion.imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
	bufferCopyRegion.imageSubresource.mipLevel = 0;
	bufferCopyRegion.imageSubresource.baseArrayLayer = 0;
	bufferCopyRegion.imageSubresource.layerCount = 1;
	bufferCopyRegion.imageExtent.width = m_Specification.Width;
	bufferCopyRegion.imageExtent.height = m_Specification.Height;
	bufferCopyRegion.imageExtent.depth = 1;
	bufferCopyRegion.bufferOffset = 0;

	// Copy mip levels from staging buffer
	vkCmdCopyBufferToImage(
		copyCmd,
		stagingBuffer,
		info.Image,
		VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
		1,
		&bufferCopyRegion);

	uint32_t mipCount = m_Specification.GenerateMips ? GetMipLevelCount() : 1;
	if (mipCount > 1) // Mips to generate
	{
		Utils::InsertImageMemoryBarrier(copyCmd, info.Image,
			VK_ACCESS_TRANSFER_WRITE_BIT, VK_ACCESS_TRANSFER_READ_BIT,
			VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
			VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
			subresourceRange);
	}
	else
	{
		Utils::InsertImageMemoryBarrier(copyCmd, info.Image,
			VK_ACCESS_TRANSFER_READ_BIT, VK_ACCESS_SHADER_READ_BIT,
			VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, image->GetDescriptorInfoVulkan().imageLayout,
			VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
			subresourceRange);
	}

	device->FlushCommandBuffer(copyCmd);

	// Clean up staging resources
	allocator.DestroyBuffer(stagingBuffer, stagingBufferAllocation);

	if (m_Specification.GenerateMips && mipCount > 1)
		GenerateMips();
}

对于Cube纹理，就是imageCreateInfo.arrayLayers = 6; // cubeMap本质是一个图片数组，arrayLayers来表示6个面，其他类似

游戏引擎开发实践（阴影篇）

Sat, 25 Oct 2025 20:49:55 +0800

在OpenGL渲染器中已经实现了硬阴影以及PCF，PCSS，在这里再引入级联阴影

级联阴影

传统阴影做法与问题

截图来自我的OpenGL延迟渲染器

1
2

float orthoSize = 10.0f;
glm::mat4 lightProjection = glm::ortho(-orthoSize, orthoSize, -orthoSize, orthoSize, near_plane, far_plane);

这行代码决定了从光源位置渲染阴影贴图的正交投影矩阵，他决定了阴影的覆盖范围，当设置为10时，硬阴影以及PCSS效果如下

当设置为100时，硬阴影与PCSS效果如下,可以看到硬阴影锯齿变得严重起来，软阴影也糊成一片了

最后设置为1000,阴影以及完全失效了，因为覆盖区域太大了，整个模型也就占几个像素，导致模型边缘区域都是一样的判断结果，另外bias也失效了

所以使用大场景时，阴影需要覆盖很大的区域，导致一大片位置只在阴影贴图中占一小部分，会导致近处模型阴影明显出问题

级联阴影的原理

首先对摄像机坐标系下的视锥体分割，分割为不同深度的分段
在光源的视角下，生成每个视锥体分段的“包围盒”，这个包围盒是生成阴影贴图时正交投影的重要参考

视锥体分段

目的是把相机的视锥体（camera frustum）沿深度方向划分成多个“级联”

常用方法叫PSSM

ds表示阴影贴图的像素长度
dp屏幕像素长度
dp/ds可以理解为阴影的锯齿误差
为了使整个画面的阴影看起来质量一致，不因到相机的距离而发生明显质量变化，应该让dp/ds 是一个常数（dp/ds：表示“一个阴影贴图像素在屏幕上看起来有多大”）
由dp/dy = n/z，得出dp = ndy/z
φ和θ分别表示曲面法线与屏幕和阴影贴图平面之间的角度。由dz/cosθ = dy/cosφ，得出dy = dzcosφ/cosθ
所以dp = ndzcosφ/zcosθ，dp/ds = ndzcosφ/zdscosθ

看看这个式子：

$$ \frac{dp}{ds} = \frac{n dz \cos\phi}{z \, ds \cos\theta} $$

它告诉你阴影质量受哪些因素影响：

因素	影响	含义
n / z	↓ 随距离增大而减小	物体越远，阴影在屏幕上越小 → 锯齿更明显
cosφ / cosθ	与法线角度相关	面倾斜时投影面积不同，阴影更“拉伸”
dz/ds	光源深度变化与贴图分辨率关系	光源投影范围越大（dz 大），每个像素越模糊
→ dp/ds 越大，阴影越锯齿

最后如果用对数拆分方案，可以推导出

其中Z是视锥体分段处的值,对应上图中的Ci，n是近平面距离，f是远平面距离，N是视锥体分割的个数，一般是1~4个

如果使用均匀拆分方案

而PSSM方法是他们的折中，用系数λ来加权平均，

float nd = camera->near;
float fd = camera->far;
float lambda = 0.75;
float ratio = fd / nd;
frustums[0].near(nd);
for(int i = 1 ; i < 分段个数 ; i++) 
{
	float si = i / (float)分段个数;
	float t_near = lambda * (nd * powf(ratio, si)) + (1 - lambda) * (nd + (fd - nd) * si);
	float far = near * 1.005f;//略微增加重合，避免断裂
	frustums[i].near(near);
	frustums[i-1].far(far);
}
frustums[分段个数-1].far(fd);

到这里就拿到了每个视锥体切片的Near和Far是多少，就是上图中的C0，C1…

光源空间中正交投影

先计算出摄像机视锥体分块在世界空间中的坐标（就是计算了近平面和远平面组成的立方体的8个顶点的世界坐标）

glm::vec3 viewPos = camera->Position;
glm::vec3 viewDir = camera->Front;
glm::vec3 up(0.0f, 1.0f, 0.0f);
glm::vec3 right = glm::cross(viewDir, up);
for(int i = 0 ; i < 分块个数 ; i++) 
{
   Frustum& frustum = m_frustums[i];
   glm::vec3 fc = viewPos + viewDir * frustum.far(); // 当前的Near
   glm::vec3 nc = viewPos + viewDir * frustum.near(); // 当前的Far
   right = glm::normalize(right);   // 相机的右方向
   up = glm::normalize(glm::cross(right, viewDir)); // 计算正交基的上方向
   // 计算当前分片的近平面和远平面宽高的一半
   float near_height = tan(frustum.fov() / 2.0f) * frustum.near();
   float near_width = near_height * frustum.ratio();
   float far_height = tan(frustum.fov() / 2.0f) * frustum.far();
   float far_width = far_height * frustum.ratio();
   //记录视锥8个顶点
   frustum.m_points[0] = nc - up * near_height - right * near_width;
   frustum.m_points[1] = nc + up * near_height - right * near_width;
   frustum.m_points[2] = nc + up * near_height + right * near_width;
   frustum.m_points[3] = nc - up * near_height + right * near_width;
   frustum.m_points[4] = fc - up * far_height - right * far_width;
   frustum.m_points[5] = fc + up * far_height - right * far_width;
   frustum.m_points[6] = fc + up * far_height + right * far_width;
   frustum.m_points[7] = fc - up * far_height + right * far_width;
}

利用分块的各顶点坐标，计算摄像机视锥体分段的“包围盒”，从而计算出分块对应的正交投影矩阵（最终结果就是每个级联块都有了自己的变换矩阵，我想把模型放在哪个级联块中，就乘以他的变换矩阵）

glm::mat4 lightProjMat;
for(int i = 0 ; i < 分块个数 ; i++) 
{
   //1. 找出光空间中八个顶点的最大最小z值（光源视角下z的范围）
   Frustum& frustum = m_frustums[i];
   glm::vec3 max(-1000.0f, -1000.0f, 0.0f);
   glm::vec3 min(1000.0f, 1000.0f, 0.0f);
   glm::vec4 transf = lightViewMat * glm::vec4(frustum.m_points[0], 1.0f);
   min.z = transf.z;
   max.z = transf.z;
   for(int j = 1 ; j < 8 ; j++) 
   {
     transf = lightViewMat * glm::vec4(frustum.m_points[j], 1.0f);
     if(transf.z > max.z) { max.z = transf.z; }
     if(transf.z < min.z) { min.z = transf.z; }
   }
//1.1 扩展光视锥体的大小，使其包括所有的遮挡物
   for(int j=0; j<场景中物体包围球个数; j++)
   {
   	transf = lightViewMat * vec4f(objBSphereCenter[j], 1.0f);
		if(transf.z + objBSphereRadius[j] > max.z)
		{
			max.z = transf.z + objBSphereRadius[j];
		}
   }
   //2. 设光空间的正交投影矩阵为ortho，他的x,y∈[-1,1]，z∈[-tmax.z,-tmin.z]
   //使用x,y∈[-1,1]，是因为每个分块的投影矩阵都可以使用单位x,y缩放平移后获得
   //使用z∈[-tmax.z,-tmin.z]，是因为摄像机空间指向负Z方向，而glm::ortho传入的是近平面和远平面指向正Z方向
   glm::mat4 ortho = glm::ortho(-1.0f, 1.0f, -1.0f, 1.0f, -tmax.z, -tmin.z);
//2.1 在光空间中，找出视锥体切片各顶点的x、y的标准设备坐标范围。因为我们设的投影矩阵的x、y都是标准设备坐标，我们需要求出x、y的变化范围，以便对ortho进行缩放平移
   glm::mat4 lightVP = ortho * lightViewMat;
   for(int j = 0 ; j < 8 ; j++) 
   {
     transf = lightVP * glm::vec4(frustum.m_points[j], 1.0f);
     transf.x /= transf.w;
     transf.y /= transf.w;
     if(transf.x > max.x) { max.x = transf.x; }
     if(transf.x < min.x) { min.x = transf.x; }
     if(transf.y > max.y) { max.y = transf.y; }
     if(transf.y < min.y) { min.y = transf.y; }
   }
//2.2 根据正交投影矩阵的公式，设置缩放平移量(计算过程在后面)
   glm::vec2 scale(2.0f / (max.x - min.x), 2.0f / (max.y - min.y));
   glm::vec2 offset(-0.5f * (max.x + min.x) * scale.x, -0.5f * (max.y + min.y) * scale.y);
   glm::mat4 crop = glm::mat4(1.0);
//2.3 设置缩放平移的变换矩阵
   crop[0][0] = scale.x;
   crop[1][1] = scale.y;
   crop[0][3] = offset.x;
   crop[1][3] = offset.y;
   crop = glm::transpose(crop);//注意glm按列储存，实际矩阵要转置
//2.4 计算出光空间中的正交投影矩阵
   lightProjMat = crop * ortho;
   //保存光空间的投影矩阵
   projection_matrices[i] = lightProjMat;
   //保存世界坐标到光空间变换的矩阵
   crop_matrices[i] = lightProjMat * lightViewMat;
}

计算出摄像机视锥体分块的远平面在摄像机空间中投影后的位置，并把他变换到[0.0，1.0]。这么做的目的是为了在中判断某个片段属于哪个分块

for(int i = 0 ; i < 分块个数 ; i++) 
{
   far_bounds[i] =0.5f*((-1.0f * frustums[i].far() * projMat[2][2] + projMat[3][2])/
   frustums[i].far())+0.5f;
 }

游戏引擎开发实践（实例化渲染）

Tue, 21 Oct 2025 20:11:16 +0800

废弃

StaticMesh收集

首先从Scene的Entity找到包含StaticMesh组件的Entity，第一步剔除不显示的Mesh

void Scene::CollectRenderableEntities(Ref<SceneRender> sceneRender)
{
	// 收集StaticMesh
	auto allEntityOwnMesh = GetAllEntitiesWith<StaticMeshComponent>();
	for (auto entity : allEntityOwnMesh) {
		auto& staticMeshComponent = allEntityOwnMesh.get<StaticMeshComponent>(entity);
		if (!staticMeshComponent.Visible) continue;
		Ref<MeshSource> mesh = AssetManager::GetAsset<MeshSource>(staticMeshComponent.StaticMesh);
		if (mesh == nullptr) continue;
		Entity e = Entity(entity, this);
		glm::mat4 transform = GetWorldSpaceTransformMatrix(e);
		sceneRender->SubmitStaticMesh(mesh, transform);
	}
}

另外因为层级结构，需要递归处理每个Mesh的模型变换矩阵

	glm::mat4 Scene::GetWorldSpaceTransformMatrix(Entity entity)
	{
		glm::mat4 transform(1.0f);

		Entity parent = GetEntityByUUID(entity.GetParentUUID());
		if (parent)
			transform = GetWorldSpaceTransformMatrix(parent);

		return transform * entity.Transform().GetTransform();
	}

最后注册这个Mesh，携带它的模型变换。

注册Mesh到StaticDrawList

遍历每个SubMesh

计算他的全局变换矩阵
获取MeshKey（用于实例化，这样做自动把一样的SubMesh自动放在了一起），这里Key就表示同一个SubMesh同一个材质，就属于实例化的一个
缓存对应的变换矩阵
增加对应DrawCommand的实例数量，最终每一个MeshKey对应的DrawCommand都会变成一个DrawCall

struct StaticDrawCommand
{
    Ref<MeshSource> MeshSource;
    uint32_t SubmeshIndex;
    Ref<MaterialAsset> MaterialAsset;
    uint32_t InstanceCount = 0;
};
void SceneRender::SubmitStaticMesh(Ref<MeshSource> meshSource, const glm::mat4& transform) {
	const auto& submeshData = meshSource->GetSubmeshes();
	for(uint32_t submeshIndex = 0; submeshIndex <submeshData.size(); submeshIndex++){
		glm::mat4 submeshTransform = transform * submeshData[submeshIndex].Transform;  // subMesh的全局变换
		AssetHandle materialHandle = meshSource->GetMaterialHandle(submeshData[submeshIndex].MaterialIndex);
		Ref<MaterialAsset> material = AssetManager::GetAsset<MaterialAsset>(materialHandle);// subMesh的材质索引;
		HZ_CORE_VERIFY(material);
		MeshKey meshKey = { meshSource->Handle, materialHandle, submeshIndex, false };
		// 缓存变换矩阵
		auto& transformStorage = m_MeshTransformMap[meshKey].Transforms.emplace_back(); // 对于每一种MeshKey，都存储多个Transforms矩阵，表示多个实例
		transformStorage.MRow[0] = { submeshTransform[0][0], submeshTransform[1][0], submeshTransform[2][0], submeshTransform[3][0] };
		transformStorage.MRow[1] = { submeshTransform[0][1], submeshTransform[1][1], submeshTransform[2][1], submeshTransform[3][1] };
		transformStorage.MRow[2] = { submeshTransform[0][2], submeshTransform[1][2], submeshTransform[2][2], submeshTransform[3][2] };
		// 缓存绘制命令
		auto& destDrawList = m_StaticMeshDrawList;
		auto& dc = destDrawList[meshKey];
		dc.MeshSource = meshSource;
		dc.SubmeshIndex = submeshIndex;
		dc.MaterialAsset = material;
		dc.InstanceCount++;
	}
};

缓存所有DrawCommand的变换矩阵

注册完成后，处理模型变换信息，这里使用的是顶点缓冲区来存储模型变换信息，下面代码作用就是把每个要渲染的SubMesh的变换矩阵按顺序存在一个顶点缓冲区，等待渲染时绑定

// 收集所有参与渲染的Mesh的变换矩阵存储在m_SubmeshTransformBuffers
{
    uint32_t frameIndex = Renderer::GetCurrentFrameIndex();

    uint32_t offset = 0;
    for (auto& [key, transformData] : m_MeshTransformMap)
    {
        transformData.TransformOffset = offset * sizeof(TransformVertexData);
        for (const auto& transform : transformData.Transforms)
        {
            m_SubmeshTransformBuffers[frameIndex].Data[offset] = transform; 
            offset++;
        }
    }
    // 在这里所有的变换矩阵已经上传到GPU（RT线程）
    m_SubmeshTransformBuffers[frameIndex].Buffer->SetData(m_SubmeshTransformBuffers[frameIndex].Data, offset * sizeof(TransformVertexData)); 
}

下面要渲染时就按顺序渲染

for (auto& [meshKey, drawCommand] : m_StaticMeshDrawList)
{
    const auto& transformData = m_MeshTransformMap.at(meshKey); // 其中已经记录好了每个DrawCall对应的变换矩阵偏移
    Renderer::RenderStaticMeshWithMaterial(m_CommandBuffer, m_GeoPipeline,drawCommand.MeshSource, drawCommand.SubmeshIndex, drawCommand.MaterialAsset->GetMaterial(), m_SubmeshTransformBuffers[frameIndex].Buffer, transformData.TransformOffset, drawCommand.InstanceCount);
}

实例化渲染

绑定整个Mesh的顶点数据
绑定所有变换矩阵的缓冲区也作为顶点数据，并设置正确的偏移量
绑定索引
材质相关

void VulkanRenderer::RenderStaticMeshWithMaterial(Ref<RenderCommandBuffer> renderCommandBuffer, Ref<Pipeline> pipeline, Ref<MeshSource> meshSource, uint32_t submeshIndex, Ref<Material> material, Ref<VertexBuffer> transformBuffer, uint32_t transformOffset, uint32_t instanceCount)
{
	HZ_CORE_ASSERT(meshSource);
	HZ_CORE_ASSERT(material);
	Ref<VulkanMaterial> vulkanMaterial = material.As<VulkanMaterial>();
	Renderer::Submit([renderCommandBuffer, pipeline,meshSource, submeshIndex, vulkanMaterial, transformBuffer, transformOffset, instanceCount]() mutable {
		uint32_t frameIndex = Renderer::RT_GetCurrentFrameIndex();
		VkCommandBuffer commandBuffer = renderCommandBuffer.As<VulkanRenderCommandBuffer>()->GetActiveCommandBuffer();
		Ref<VulkanVertexBuffer> meshVertBuffer = meshSource->GetVertexBuffer().As<VulkanVertexBuffer>();
		VkBuffer vkMeshVertBuffer = meshVertBuffer->GetVulkanBuffer();
		VkDeviceSize vertexOffsets[1] = { 0 };
		vkCmdBindVertexBuffers(commandBuffer, 0, 1, &vkMeshVertBuffer, vertexOffsets);  // 把整个Mesh的顶点都绑定
		Ref<VulkanVertexBuffer> vulkanTransformBuffer = transformBuffer.As<VulkanVertexBuffer>();
		VkBuffer vkTransformBuffer = vulkanTransformBuffer->GetVulkanBuffer();
		VkDeviceSize instanceOffsets[1] = { transformOffset };
		vkCmdBindVertexBuffers(commandBuffer, 1, 1, &vkTransformBuffer, instanceOffsets); // 第二个顶点缓冲区绑定当前SubMesh的变换矩阵数据

		auto vulkanMeshIB = Ref<VulkanIndexBuffer>(meshSource->GetIndexBuffer());
		VkBuffer ibBuffer = vulkanMeshIB->GetVulkanBuffer();
		vkCmdBindIndexBuffer(commandBuffer, ibBuffer, 0, VK_INDEX_TYPE_UINT32); // 索引缓冲区全绑定

		// 每个材质绑定自己的 Set=1
		VkDescriptorSet matSet = vulkanMaterial->GetDescriptorSets()[frameIndex];
		vkCmdBindDescriptorSets(commandBuffer, VK_PIPELINE_BIND_POINT_GRAPHICS,
			pipeline.As<VulkanPipeline>()->GetVulkanPipelineLayout(),
			1, // Set=1
			1, &matSet,
			0, nullptr);
		const auto& submeshes = meshSource->GetSubmeshes();
		const auto& submesh = submeshes[submeshIndex];
		vkCmdDrawIndexed(commandBuffer, submesh.IndexCount/*索引数量*/, instanceCount/*实例数量*/, submesh.BaseIndex/*索引缓冲区的偏移*/, submesh.BaseVertex/*顶点偏移*/, 0/*实例化ID开始的编号*/);
		
	});
}

另外要注意实例的变换矩阵的切换时机要在Pipeline创建时设置好,下面可以看到实例顶点布局的inputRate是设置的VK_VERTEX_INPUT_RATE_INSTANCE，这样他就会每个实例切换一次


    VkVertexInputBindingDescription& vertexInputBinding = vertexInputBindingDescriptions.emplace_back();
    vertexInputBinding.binding = 0;
    vertexInputBinding.stride = vertexLayout.GetStride();
    vertexInputBinding.inputRate = VK_VERTEX_INPUT_RATE_VERTEX;

    if (instanceLayout.GetElementCount())
    {
        VkVertexInputBindingDescription& instanceInputBinding = vertexInputBindingDescriptions.emplace_back();
        instanceInputBinding.binding = 1;
        instanceInputBinding.stride = instanceLayout.GetStride();
        instanceInputBinding.inputRate = VK_VERTEX_INPUT_RATE_INSTANCE;
    }

Shader

#version 450 core
#ifdef VERTEX_SHADER
layout(binding = 0) uniform UniformBufferObject {
    mat4 view;
    mat4 proj;
	float width;
	float height;
	float Near;
	float Far;
} ubo;

layout(location = 0) in vec3 inPosition;
layout(location = 1) in vec3 inNormal;
layout(location = 2) in vec3 inTangent;
layout(location = 3) in vec3 inBinormal; 
layout(location = 4) in vec2 inTexCoord;
layout(location = 5) in vec4 a_MRow0;
layout(location = 6) in vec4 a_MRow1;
layout(location = 7) in vec4 a_MRow2;

layout(location = 0) out vec3 fragWorldPos;
layout(location = 1) out vec3 fragWorldNormal;
layout(location = 2) out vec3 fragWorldTangent;
layout(location = 3) out vec3 fragWorldBinormal;
layout(location = 4) out vec2 fragTexCoord;

void main() {
    mat4 transform = mat4(
		vec4(a_MRow0.x, a_MRow1.x, a_MRow2.x, 0.0),
		vec4(a_MRow0.y, a_MRow1.y, a_MRow2.y, 0.0),
		vec4(a_MRow0.z, a_MRow1.z, a_MRow2.z, 0.0),
		vec4(a_MRow0.w, a_MRow1.w, a_MRow2.w, 1.0)
	);
    vec4 worldPos = transform * vec4(inPosition, 1.0);
    fragWorldPos = worldPos.xyz;

    mat3 modelRot = mat3(transform);
    fragWorldNormal = normalize(modelRot * inNormal);
    fragWorldTangent = normalize(modelRot * inTangent);
    fragWorldBinormal = normalize(modelRot * inBinormal);
    fragTexCoord = inTexCoord;
    gl_Position = ubo.proj * ubo.view * worldPos;
}
#endif

RenderDoc可以看到，只有一个DrawCall

游戏引擎开发实践（动画篇）

Mon, 20 Oct 2025 20:17:44 +0800

骨骼构建

骨骼信息读取

骨骼数据包含

骨骼名称
每个骨骼的父骨骼索引
每个骨骼相对于父骨骼的变换（平移、旋转、缩放）
整个骨架的变换矩阵

	std::vector<std::string> m_BoneNames;
	std::vector<uint32_t> m_ParentBoneIndices;
	// rest pose of skeleton. All in bone-local space (i.e. translation/rotation/scale relative to parent)
	std::vector<glm::vec3> m_BoneTranslations;
	std::vector<glm::quat> m_BoneRotations;
	std::vector<glm::vec3> m_BoneScales;
	// The skeleton itself can have a transform
	// Notably this happens if the whole "armature" is rotated or scaled in DCC tool
	glm::mat4 m_Transform;

另外骨骼信息应该独属于Mesh，应该用unique_ptr包裹指针

第一步直接把所以SubMesh的骨骼信息提取到Vector，这一步只是把Name提取出来了

void BoneHierarchy::ExtractBones()
{
	// Note: ASSIMP does not appear to support import of digital content files that contain _only_ an armature/skeleton and no mesh.
	for (uint32_t meshIndex = 0; meshIndex < m_Scene->mNumMeshes; ++meshIndex)
	{
		const aiMesh* mesh = m_Scene->mMeshes[meshIndex];
		for (uint32_t boneIndex = 0; boneIndex < mesh->mNumBones; ++boneIndex)
		{
			m_Bones.emplace(mesh->mBones[boneIndex]->mName.C_Str());
		}
	}

	// Extract also any nodes that are animated (but don't have any skin bound to them)
	for (uint32_t animationIndex = 0; animationIndex < m_Scene->mNumAnimations; ++animationIndex)
	{
		const aiAnimation* animation = m_Scene->mAnimations[animationIndex];
		for (uint32_t channelIndex = 0; channelIndex < animation->mNumChannels; ++channelIndex)
		{
			const aiNodeAnim* nodeAnim = animation->mChannels[channelIndex];
			m_Bones.emplace(nodeAnim->mNodeName.C_Str());
		}
	}
}

第二步一套代码从assimp中读取了骨骼信息

void BoneHierarchy::TraverseNode(aiNode* node, Skeleton* skeleton, const glm::mat4& parentTransform)
{
	if (m_Bones.find(node->mName.C_Str()) != m_Bones.end())
	{
		skeleton->SetTransform(parentTransform);
		aiNode* parent = node->mParent;
		while (parent)
		{
			parent->mTransformation = aiMatrix4x4();
			parent = parent->mParent;
		}
		TraverseBone(node, skeleton, Skeleton::NullIndex);
	}
	else
	{
		auto transform = parentTransform * Utils::Mat4FromAIMatrix4x4(node->mTransformation);
		for (uint32_t nodeIndex = 0; nodeIndex < node->mNumChildren; ++nodeIndex)
		{
			TraverseNode(node->mChildren[nodeIndex], skeleton, transform);
		}
	}
}

void BoneHierarchy::TraverseBone(aiNode* node, Skeleton* skeleton, uint32_t parentIndex)
{
	uint32_t boneIndex = skeleton->AddBone(node->mName.C_Str(), parentIndex, Utils::Mat4FromAIMatrix4x4(node->mTransformation));
	for (uint32_t nodeIndex = 0; nodeIndex < node->mNumChildren; ++nodeIndex)
	{
		if (m_Bones.find(node->mChildren[nodeIndex]->mName.C_Str()) != m_Bones.end())
		{
			TraverseBone(node->mChildren[nodeIndex], skeleton, boneIndex);
		}
		else
		{
			// do not traverse any further.
			// It is not supported to have a non-bone and then more bones below it.
		}
	}
}

蒙皮权重处理

首先定义骨骼权重信息

	struct BoneInfluence
	{
		uint32_t BoneInfoIndices[4] = { 0, 0, 0, 0 };  // 影响当前顶点的骨骼索引（最多4个）
		float Weights[4] = { 0.0f, 0.0f, 0.0f, 0.0f }; // 对应骨骼的影响权重（最多4个）

		void AddBoneData(uint32_t boneInfoIndex, float weight)
		{
			if (weight < 0.0f || weight > 1.0f)
			{
				HZ_CORE_WARN("Vertex bone weight is out of range. We will clamp it to [0, 1] (BoneID={0}, Weight={1})", boneInfoIndex, weight);
				weight = std::clamp(weight, 0.0f, 1.0f);
			}
			if (weight > 0.0f)
			{
				for (size_t i = 0; i < 4; i++)
				{
					if (Weights[i] == 0.0f)
					{
						BoneInfoIndices[i] = boneInfoIndex;
						Weights[i] = weight;
						return;
					}
				}

				// Note: when importing from assimp we are passing aiProcess_LimitBoneWeights which automatically keeps only the top N (where N defaults to 4)
				//       bone weights (and normalizes the sum to 1), which is exactly what we want.
				//       So, we should never get here.
				HZ_CORE_WARN("Vertex has more than four bones affecting it, extra bone influences will be discarded (BoneID={0}, Weight={1})", boneInfoIndex, weight);
			}
		}

		void NormalizeWeights()
		{
			float sumWeights = 0.0f;
			for (size_t i = 0; i < 4; i++)
			{
				sumWeights += Weights[i];
			}
			if (sumWeights > 0.0f)
			{
				for (size_t i = 0; i < 4; i++)
				{
					Weights[i] /= sumWeights;
				}
			}
		}
	};

另外还要骨骼信息

struct BoneInfo
{
	glm::mat4 InverseBindPose;
	uint32_t BoneIndex;
	BoneInfo(const glm::mat4& inverseBindPose, uint32_t boneIndex)
		: InverseBindPose(inverseBindPose), BoneIndex(boneIndex) {
	}
};

下来遍历每个SubMesh的骨骼信息来处理每个顶点的权重信息

				// skinning weights
		meshSource->m_Skeleton = AssimpAnimationImporter::ImportSkeleton(scene);
		if (meshSource->HasSkeleton())
		{
			HZ_CORE_INFO_TAG("Mesh", "开始处理骨骼信息");
			meshSource->m_BoneInfluences.resize(meshSource->m_Vertices.size());
			for (uint32_t m = 0; m < scene->mNumMeshes; m++)
			{
				aiMesh* mesh = scene->mMeshes[m];
				Submesh& submesh = meshSource->m_Submeshes[m];

				if (mesh->mNumBones > 0)
				{
					submesh.IsRigged = true;
					for (uint32_t i = 0; i < mesh->mNumBones; i++)
					{
						aiBone* bone = mesh->mBones[i];
						bool hasNonZeroWeight = false;
						for (size_t j = 0; j < bone->mNumWeights; j++)
						{
							if (bone->mWeights[j].mWeight > 0.000001f)
							{
								hasNonZeroWeight = true;
								break;
							}
						}
						if (!hasNonZeroWeight)
							continue;

						// Find bone in skeleton
						uint32_t boneIndex = meshSource->m_Skeleton->GetBoneIndex(bone->mName.C_Str());
						if (boneIndex == Skeleton::NullIndex)
						{
							HZ_CORE_ERROR_TAG("Animation", "Could not find mesh bone '{}' in skeleton!", bone->mName.C_Str());
						}

						uint32_t boneInfoIndex = ~0;
						for (size_t j = 0; j < meshSource->m_BoneInfo.size(); ++j)
						{
							if (meshSource->m_BoneInfo[j].BoneIndex == boneIndex)
							{
								boneInfoIndex = static_cast<uint32_t>(j);
								break;
							}
						}
						if (boneInfoIndex == ~0)
						{
							boneInfoIndex = static_cast<uint32_t>(meshSource->m_BoneInfo.size());
							const auto& boneInfo = meshSource->m_BoneInfo.emplace_back(Utils::Mat4FromAIMatrix4x4(bone->mOffsetMatrix), boneIndex);
							HZ_CORE_INFO_TAG("Mesh", "BoneInfo for bone '{0}'", bone->mName.C_Str());
							HZ_CORE_INFO_TAG("Mesh", "  SubMeshIndex = {0}", m);
							HZ_CORE_INFO_TAG("Mesh", "  BoneIndex = {0}", boneIndex);
							glm::vec3 translation;
							glm::quat rotationQuat;
							glm::vec3 scale;
							Math::DecomposeTransform(boneInfo.InverseBindPose, translation, rotationQuat, scale);
							glm::vec3 rotation = glm::degrees(glm::eulerAngles(rotationQuat));
							HZ_CORE_INFO_TAG("Mesh", "  Inverse Bind Pose = {");
							HZ_CORE_INFO("    translation: ({0:8.4f}, {1:8.4f}, {2:8.4f})", translation.x, translation.y, translation.z);
							HZ_CORE_INFO("    rotation:    ({0:8.4f}, {1:8.4f}, {2:8.4f})", rotation.x, rotation.y, rotation.z);
							HZ_CORE_INFO("    scale:       ({0:8.4f}, {1:8.4f}, {2:8.4f})", scale.x, scale.y, scale.z);
							HZ_CORE_INFO("  }");
						}

						for (size_t j = 0; j < bone->mNumWeights; j++)
						{
							int VertexID = submesh.BaseVertex + bone->mWeights[j].mVertexId;
							float Weight = bone->mWeights[j].mWeight;
							meshSource->m_BoneInfluences[VertexID].AddBoneData(boneInfoIndex, Weight);
						}
					}
				}
			}

			for (auto& boneInfluence : meshSource->m_BoneInfluences)
			{
				boneInfluence.NormalizeWeights();
			}
		}

至此为MeshSource导入了骨骼信息，封装为

1
2

	std::vector<BoneInfluence> m_BoneInfluences; // 每个顶点对应的骨骼影响数据
	std::vector<BoneInfo> m_BoneInfo; // 骨骼信息

动画构建

GLTF的动画结构

采样器:定义了有几个关键帧，以及每个关键帧对应的骨骼运动信息
通道：指明一根骨骼由哪个采样器控制，以及控制方式（平移、旋转、缩放）

使用游戏引擎制作动画时的关键帧是统一设置了所有骨骼的变换信息，GLTF对每个骨骼单独控制，节省了很多空间，因为可能两个关键帧之间某些骨骼并没有变换，但是统一存储还得存一遍。下面是一个开火动画的例子

 "animations" : [    
        {
            "name" : "Fire", // 开火动画的例子
            "samplers" : [
                {
                    "input" : 7, // 存储的是所有关键帧的时间点（7只是访问accessors 数组的索引）
                    "interpolation" : "LINEAR", // 采用线性插值
                    "output" : 8 // 与input里一一对应的关键帧的数据
                },
                {
                    "input" : 7,
                    "interpolation" : "LINEAR",
                    "output" : 9
                }
            ]，
            "channels" : [
                {
                    "sampler" : 0, // 指定一个采样器
                    "target" : {
                        "node" : 5, // 指定这个采样器作用哪个骨骼
                        "path" : "translation" // 作用的方式
                    }
                },
                {
                    "sampler" : 1,
                    "target" : {
                        "node" : 5,
                        "path" : "rotation"
                    }
                }
            ]
        }
    ]

动画数据本质就是存储每个关键帧的骨骼变换数据，GLTF这样存储是一种优化手段

GLTF动画数据导入

定义关键帧

template<typename T> struct KeyFrame
{
	float FrameTime;
	T Value;
	KeyFrame(const float frameTime, const T& value) : FrameTime(frameTime), Value(value) {}
};

定义Channel结构,对于一组关键帧，存储一根骨骼每个关键帧的变换数据，以及骨骼的Index（捕捉了 “单根骨骼的完整动画轨迹”）

struct Channel
{
	std::vector<KeyFrame<glm::vec3>> Translations;
	std::vector<KeyFrame<glm::quat>> Rotations;
	std::vector<KeyFrame<glm::vec3>> Scales;
	uint32_t Index;
};

下面给出如何封装所有骨骼的Channel

static std::vector<Channel> ImportChannels(aiAnimation* anim, const Skeleton& skeleton, const bool isMaskedRootMotion, const glm::vec3& rootTranslationMask, float rootRotationMask)
{
	std::vector<Channel> channels;

	std::unordered_map<std::string_view, uint32_t> boneIndices;
	std::unordered_set<uint32_t> rootBoneIndices;
	for (uint32_t i = 0; i < skeleton.GetNumBones(); ++i)
	{
		boneIndices.emplace(skeleton.GetBoneName(i), i + 1);  // 0 is reserved for root motion channel    boneIndices are base=1
		if (skeleton.GetParentBoneIndex(i) == Skeleton::NullIndex)
			rootBoneIndices.emplace(i + 1);
	}

	std::map<uint32_t, aiNodeAnim*> validChannels;
	for (uint32_t channelIndex = 0; channelIndex < anim->mNumChannels; ++channelIndex)
	{
		aiNodeAnim* nodeAnim = anim->mChannels[channelIndex];
		auto it = boneIndices.find(nodeAnim->mNodeName.C_Str());
		if (it != boneIndices.end())
		{
			validChannels.emplace(it->second, nodeAnim);   // validChannels.first is base=1,  .second is node pointer
		}
	}

	channels.resize(skeleton.GetNumBones() + 1);   // channels is base=1

	// channels don't necessarily have first frame at time zero.
	// We can just generate a dummy key frame, but that can end up looking a bit odd (in particular,
	// looping animations appear to pause for a split second each time they loop and encounter the
	// dummy key frame.
	// So instead, we clip the animation time horizon.
	double firstFrameDelta = DBL_MAX;
	double animationDuration = anim->mDuration;
	for (uint32_t boneIndex = 1; boneIndex < channels.size(); ++boneIndex)
	{
		if (auto validChannel = validChannels.find(boneIndex); validChannel != validChannels.end())
		{
			auto nodeAnim = validChannel->second;
			if (nodeAnim->mNumPositionKeys > 0)
				firstFrameDelta = std::min(firstFrameDelta, nodeAnim->mPositionKeys[0].mTime);

			if (nodeAnim->mNumRotationKeys > 0)
				firstFrameDelta = std::min(firstFrameDelta, nodeAnim->mRotationKeys[0].mTime);

			if (nodeAnim->mNumScalingKeys > 0)
				firstFrameDelta = std::min(firstFrameDelta, nodeAnim->mScalingKeys[0].mTime);
		}
	}

	anim->mDuration -= firstFrameDelta;

	// The rest of the code assumes non-zero animation duration.
	// Enforce that here.
	if (anim->mDuration <= 0.0) {
		anim->mDuration = 1.0;
	}

	for (uint32_t boneIndex = 1; boneIndex < channels.size(); ++boneIndex)
	{
		Channel& channel = channels[boneIndex];
		channel.Index = boneIndex;
		if (auto validChannel = validChannels.find(boneIndex); validChannel != validChannels.end())
		{
			auto nodeAnim = validChannel->second;
			channel.Translations.reserve(nodeAnim->mNumPositionKeys + 2); // +2 because worst case we insert two more keys
			channel.Rotations.reserve(nodeAnim->mNumRotationKeys + 2);
			channel.Scales.reserve(nodeAnim->mNumScalingKeys + 2);
			// Note: There is no need to check for duplicate keys (i.e. multiple keys all at same frame time)
			//       because Assimp throws these out for us
			for (uint32_t keyIndex = 0; keyIndex < nodeAnim->mNumPositionKeys; ++keyIndex)
			{
				aiVectorKey key = nodeAnim->mPositionKeys[keyIndex];
				float frameTime = std::clamp(static_cast<float>((key.mTime - firstFrameDelta) / anim->mDuration), 0.0f, 1.0f);
				if ((keyIndex == 0) && (frameTime > 0.0f))
				{
					channels[boneIndex].Translations.emplace_back(0.0f, glm::vec3{ static_cast<float>(key.mValue.x), static_cast<float>(key.mValue.y), static_cast<float>(key.mValue.z) });
				}
				channel.Translations.emplace_back(frameTime, glm::vec3{ static_cast<float>(key.mValue.x), static_cast<float>(key.mValue.y), static_cast<float>(key.mValue.z) });
			}
			if (channel.Translations.empty())
			{
				HZ_CORE_WARN_TAG("Animation", "No translation track found for bone '{}'", skeleton.GetBoneName(boneIndex - 1));
				channel.Translations = { {0.0f, glm::vec3{0.0f}}, {1.0f, glm::vec3{0.0f}} };
			}
			else if (channel.Translations.back().FrameTime < 1.0f)
			{
				channel.Translations.emplace_back(1.0f, channel.Translations.back().Value);
			}
			for (uint32_t keyIndex = 0; keyIndex < nodeAnim->mNumRotationKeys; ++keyIndex)
			{
				aiQuatKey key = nodeAnim->mRotationKeys[keyIndex];
				float frameTime = std::clamp(static_cast<float>((key.mTime - firstFrameDelta) / anim->mDuration), 0.0f, 1.0f);
				// WARNING: constructor parameter order for a quat is still WXYZ even if you have defined GLM_FORCE_QUAT_DATA_XYZW
				if ((keyIndex == 0) && (frameTime > 0.0f))
				{
					channel.Rotations.emplace_back(0.0f, glm::quat{ static_cast<float>(key.mValue.w), static_cast<float>(key.mValue.x), static_cast<float>(key.mValue.y), static_cast<float>(key.mValue.z) });
				}
				channel.Rotations.emplace_back(frameTime, glm::quat{ static_cast<float>(key.mValue.w), static_cast<float>(key.mValue.x), static_cast<float>(key.mValue.y), static_cast<float>(key.mValue.z) });
				HZ_CORE_ASSERT(fabs(glm::length(channels[boneIndex].Rotations.back().Value) - 1.0f) < 0.00001f);   // check rotations are normalized (I think assimp ensures this, but not 100% sure)
			}
			if (channel.Rotations.empty())
			{
				HZ_CORE_WARN_TAG("Animation", "No rotation track found for bone '{}'", skeleton.GetBoneName(boneIndex - 1));
				channel.Rotations = { {0.0f, glm::quat{1.0f, 0.0f, 0.0f, 0.0f}}, {1.0f, glm::quat{1.0f, 0.0f, 0.0f, 0.0f}} };
			}
			else if (channel.Rotations.back().FrameTime < 1.0f)
			{
				channel.Rotations.emplace_back(1.0f, channel.Rotations.back().Value);
			}
			for (uint32_t keyIndex = 0; keyIndex < nodeAnim->mNumScalingKeys; ++keyIndex)
			{
				aiVectorKey key = nodeAnim->mScalingKeys[keyIndex];
				float frameTime = std::clamp(static_cast<float>((key.mTime - firstFrameDelta) / anim->mDuration), 0.0f, 1.0f);
				if (keyIndex == 0 && frameTime > 0.0f)
				{
					channel.Scales.emplace_back(0.0f, glm::vec3{ static_cast<float>(key.mValue.x), static_cast<float>(key.mValue.y), static_cast<float>(key.mValue.z) });
				}
				channel.Scales.emplace_back(frameTime, glm::vec3{ static_cast<float>(key.mValue.x), static_cast<float>(key.mValue.y), static_cast<float>(key.mValue.z) });
			}
			if (channel.Scales.empty())
			{
				HZ_CORE_WARN_TAG("Animation", "No scale track found for bone '{}'", skeleton.GetBoneName(boneIndex - 1));
				channel.Scales = { {0.0f, glm::vec3{1.0f}}, {1.0f, glm::vec3{1.0f}} };
			}
			else if (channel.Scales.back().FrameTime < 1.0f)
			{
				channel.Scales.emplace_back(1.0f, channels[boneIndex].Scales.back().Value);
			}
		}
		else
		{
			HZ_CORE_WARN_TAG("Animation", "No animation tracks found for bone '{}'", skeleton.GetBoneName(boneIndex - 1));

			auto translation = skeleton.GetBoneTranslations().at(boneIndex - 1);
			auto rotation = skeleton.GetBoneRotations().at(boneIndex - 1);
			auto scale = skeleton.GetBoneScales().at(boneIndex - 1);

			channel.Translations = { {0.0f, translation}, {1.0f, translation} };
			channel.Rotations = { {0.0f, rotation}, {1.0f, rotation} };
			channel.Scales = { {0.0f, scale}, {1.0f, scale} };
		}
	}

	// Create root motion channel.
	// If isMaskedRootMotion is true, then root motion channel is created by filtering components of the first channel.
	// Otherwise root motion channel is copied as-is from the first channel.
	//
	// Root motion is then removed from all "root" channels (so it doesn't get applied twice)

	HZ_CORE_ASSERT(!rootBoneIndices.empty()); // Can't see how this would ever be false!
	HZ_CORE_ASSERT(rootBoneIndices.find(1) != rootBoneIndices.end()); // First bone must be a root!

	Channel& root = channels[0];
	root.Index = 0;
	if (isMaskedRootMotion)
	{
		for (auto& translation : channels[1].Translations)
		{
			root.Translations.emplace_back(translation.FrameTime, translation.Value * rootTranslationMask);
			translation.Value *= (glm::vec3(1.0f) - rootTranslationMask);
			translation.Value += root.Translations.front().Value;
		}
		for (auto& rotation : channels[1].Rotations)
		{
			if (rootRotationMask > 0.0f)
			{
				auto angleY = Utils::AngleAroundYAxis(rotation.Value);
				root.Rotations.emplace_back(rotation.FrameTime, glm::quat{ glm::cos(angleY * 0.5f), glm::vec3{0.0f, 1.0f, 0.0f} *glm::sin(angleY * 0.5f) });
				rotation.Value = glm::conjugate(glm::quat(glm::cos(angleY * 0.5f), glm::vec3{ 0.0f, 1.0f, 0.0f } *glm::sin(angleY * 0.5f))) * rotation.Value;
				rotation.Value *= root.Rotations.front().Value;
			}
			else
			{
				root.Rotations.emplace_back(rotation.FrameTime, glm::quat{ 1.0f, 0.0f, 0.0f, 0.0f });
			}
		}
	}
	else
	{
		root.Translations = channels[1].Translations;
		root.Rotations = channels[1].Rotations;
		channels[1].Translations = { {0.0f, root.Translations.front().Value}, {1.0f, root.Translations.front().Value} };
		channels[1].Rotations = { {0.0f, root.Rotations.front().Value}, {1.0f, root.Rotations.front().Value} };
	}
	root.Scales = { {0.0f, glm::vec3{1.0f}}, {1.0f, glm::vec3{1.0f}} };

	// It is possible that there is more than one "root" bone in the asset.
	// We need to remove the root motion from all of them (otherwise those bones will move twice as fast when root motion is applied)
	for (const auto rootBoneIndex : rootBoneIndices)
	{
		// we already removed root motion from the first bone, above
		if (rootBoneIndex != 1)
		{
			for (auto& translation : channels[rootBoneIndex].Translations)
			{
				// sample root channel at this translation's frametime
				for (size_t rootFrame = 0; rootFrame < root.Translations.size() - 1; ++rootFrame)
				{
					if (root.Translations[rootFrame + 1].FrameTime >= translation.FrameTime)
					{
						const float alpha = (translation.FrameTime - root.Translations[rootFrame].FrameTime) / (root.Translations[rootFrame + 1].FrameTime - root.Translations[rootFrame].FrameTime);
						translation.Value -= glm::mix(root.Translations[rootFrame].Value, root.Translations[rootFrame + 1].Value, alpha);
						translation.Value += root.Translations.front().Value;
						break;
					}
				}
			}

			for (auto& rotation : channels[rootBoneIndex].Rotations)
			{
				// sample root channel at the this rotation's frametime
				for (size_t rootFrame = 0; rootFrame < root.Rotations.size() - 1; ++rootFrame)
				{
					if (root.Rotations[rootFrame + 1].FrameTime >= rotation.FrameTime)
					{
						const float alpha = (rotation.FrameTime - root.Rotations[rootFrame].FrameTime) / (root.Rotations[rootFrame + 1].FrameTime - root.Rotations[rootFrame].FrameTime);
						rotation.Value = glm::normalize(glm::conjugate(glm::slerp(root.Rotations[rootFrame].Value, root.Rotations[rootFrame + 1].Value, alpha)) * rotation.Value);
						rotation.Value *= root.Rotations.front().Value;
						break;
					}
				}
			}
		}
	}
	return channels;
}

压缩算法（如 ACL）要求所有骨骼的关键帧数量和时间戳严格对齐，否则无法批量处理。但是GLTF导入的骨骼关键帧是不同的采样器得到的，不一定一致，所以需要统一化

	void SanitizeChannels(std::vector<Channel>& channels)
	{
		uint32_t maxNumFrames = 2; // The rest of the code requires each channel to have at least 2 frames
		for (const auto& channel : channels)
		{
			maxNumFrames = std::max(maxNumFrames, static_cast<uint32_t>(channel.Translations.size()));
			maxNumFrames = std::max(maxNumFrames, static_cast<uint32_t>(channel.Rotations.size()));
			maxNumFrames = std::max(maxNumFrames, static_cast<uint32_t>(channel.Scales.size()));
		}

		float frameInterval = 1.0f / (maxNumFrames - 1);

		// loop over all channels and change them so they all have maxNumFrames frames.
		// add new frames where necessary by interpolating between existing frames.
		for (auto& channel : channels)
		{
			Channel newChannel;

			uint32_t translationIndex = 1;
			newChannel.Translations.reserve(maxNumFrames);
			newChannel.Translations.emplace_back(channel.Translations[0]);
			for (uint32_t i = 1; i < maxNumFrames - 1; ++i)
			{
				float frameTime = i * frameInterval;
				while ((translationIndex < channel.Translations.size()) && (channel.Translations[translationIndex].FrameTime < frameTime))
				{
					++translationIndex;
				}
				const float t = (frameTime - channel.Translations[translationIndex - 1].FrameTime) / (channel.Translations[translationIndex].FrameTime - channel.Translations[translationIndex - 1].FrameTime);
				newChannel.Translations.emplace_back(frameTime, glm::mix(channel.Translations[translationIndex - 1].Value, channel.Translations[translationIndex].Value, t));
			}
			newChannel.Translations.emplace_back(channel.Translations.back());

			uint32_t rotationIndex = 1;
			newChannel.Rotations.reserve(maxNumFrames);
			newChannel.Rotations.emplace_back(channel.Rotations[0]);
			for (uint32_t i = 1; i < maxNumFrames - 1; ++i)
			{
				float frameTime = i * frameInterval;
				while ((rotationIndex < channel.Rotations.size()) && (channel.Rotations[rotationIndex].FrameTime < frameTime))
				{
					++rotationIndex;
				}
				const float t = (frameTime - channel.Rotations[rotationIndex - 1].FrameTime) / (channel.Rotations[rotationIndex].FrameTime - channel.Rotations[rotationIndex - 1].FrameTime);
				newChannel.Rotations.emplace_back(frameTime, glm::slerp(channel.Rotations[rotationIndex - 1].Value, channel.Rotations[rotationIndex].Value, t));
			}
			newChannel.Rotations.emplace_back(channel.Rotations.back());

			uint32_t scaleIndex = 1;
			newChannel.Scales.reserve(maxNumFrames);
			newChannel.Scales.emplace_back(channel.Scales[0]);
			for (uint32_t i = 1; i < maxNumFrames - 1; ++i)
			{
				float frameTime = i * frameInterval;
				while ((scaleIndex < channel.Scales.size()) && (channel.Scales[scaleIndex].FrameTime < frameTime))
				{
					++scaleIndex;
				}
				const float t = (frameTime - channel.Scales[scaleIndex - 1].FrameTime) / (channel.Scales[scaleIndex].FrameTime - channel.Scales[scaleIndex - 1].FrameTime);
				newChannel.Scales.emplace_back(frameTime, glm::mix(channel.Scales[scaleIndex - 1].Value, channel.Scales[scaleIndex].Value, t));
			}
			newChannel.Scales.emplace_back(channel.Scales.back());

			channel.Translations = std::move(newChannel.Translations);
			channel.Rotations = std::move(newChannel.Rotations);
			channel.Scales = std::move(newChannel.Scales);
		}
	}

进一步使用acl压缩动画数据

	acl::error_result CompressChannels(const std::vector<Channel>& channels, const float fps, const Skeleton& skeleton, acl::compressed_tracks*& outCompressedTracks)
	{
		acl::iallocator& allocator = Utils::GetAnimationAllocator();
		uint32_t numTracks = static_cast<uint32_t>(channels.size());
		uint32_t numSamples = static_cast<uint32_t>(channels[0].Translations.size());
		acl::track_array_qvvf rawTrackList(allocator, numTracks);

		for (uint32_t i = 0; i < numTracks; ++i)
		{
			acl::track_desc_transformf desc;
			desc.output_index = i;
			desc.parent_index = (i == 0) ? acl::k_invalid_track_index : skeleton.GetParentBoneIndex(i - 1) + 1;  // 0 is root motion channel, 1..numBones are in the skeleton
			desc.precision = 0.0001f;
			desc.shell_distance = 3.0f;

			acl::track_qvvf rawTrack = acl::track_qvvf::make_reserve(desc, allocator, numSamples, fps);
			for (uint32_t j = 0; j < numSamples; ++j)
			{
				const auto& translation = channels[i].Translations[j].Value;
				const auto& rotation = channels[i].Rotations[j].Value;
				const auto& scale = channels[i].Scales[j].Value;

				rawTrack[j].rotation = rtm::quat_set(rotation.x, rotation.y, rotation.z, rotation.w);
				rawTrack[j].translation = rtm::vector_set(translation.x, translation.y, translation.z);
				rawTrack[j].scale = rtm::vector_set(scale.x, scale.y, scale.z);
			}
			rawTrackList[i] = std::move(rawTrack);
		}

		acl::pre_process_settings_t preProcessSettings;
		preProcessSettings.actions = acl::pre_process_actions::recommended;
		preProcessSettings.precision_policy = acl::pre_process_precision_policy::lossy;
		acl::qvvf_transform_error_metric error_metric;
		preProcessSettings.error_metric = &error_metric;
		acl::error_result result = acl::pre_process_track_list(allocator, preProcessSettings, rawTrackList);
		if (result.any())
		{
			return result;
		}

		acl::compression_settings compressSettings = acl::get_default_compression_settings();
		compressSettings.error_metric = &error_metric;
		acl::output_stats stats{ acl::stat_logging::none };
		result = acl::compress_track_list(allocator, rawTrackList, compressSettings, outCompressedTracks, stats);

		return result;
	}

ACL动画压缩

acl::error_result CompressChannels(const std::vector<Channel>& channels, const float fps, const Skeleton& skeleton, acl::compressed_tracks*& outCompressedTracks)
{
	acl::iallocator& allocator = Utils::GetAnimationAllocator();
	uint32_t numTracks = static_cast<uint32_t>(channels.size());
	uint32_t numSamples = static_cast<uint32_t>(channels[0].Translations.size());
	acl::track_array_qvvf rawTrackList(allocator, numTracks);

	for (uint32_t i = 0; i < numTracks; ++i)
	{
		acl::track_desc_transformf desc;
		desc.output_index = i;
		desc.parent_index = (i == 0) ? acl::k_invalid_track_index : skeleton.GetParentBoneIndex(i - 1) + 1;  // 0 is root motion channel, 1..numBones are in the skeleton
		desc.precision = 0.0001f;
		desc.shell_distance = 3.0f;

		acl::track_qvvf rawTrack = acl::track_qvvf::make_reserve(desc, allocator, numSamples, fps);
		for (uint32_t j = 0; j < numSamples; ++j)
		{
			const auto& translation = channels[i].Translations[j].Value;
			const auto& rotation = channels[i].Rotations[j].Value;
			const auto& scale = channels[i].Scales[j].Value;

			rawTrack[j].rotation = rtm::quat_set(rotation.x, rotation.y, rotation.z, rotation.w);
			rawTrack[j].translation = rtm::vector_set(translation.x, translation.y, translation.z);
			rawTrack[j].scale = rtm::vector_set(scale.x, scale.y, scale.z);
		}
		rawTrackList[i] = std::move(rawTrack);
	}

	acl::pre_process_settings_t preProcessSettings;
	preProcessSettings.actions = acl::pre_process_actions::recommended;
	preProcessSettings.precision_policy = acl::pre_process_precision_policy::lossy;
	acl::qvvf_transform_error_metric error_metric;
	preProcessSettings.error_metric = &error_metric;
	acl::error_result result = acl::pre_process_track_list(allocator, preProcessSettings, rawTrackList);
	if (result.any())
	{
		return result;
	}

	acl::compression_settings compressSettings = acl::get_default_compression_settings();
	compressSettings.error_metric = &error_metric;
	acl::output_stats stats{ acl::stat_logging::none };
	result = acl::compress_track_list(allocator, rawTrackList, compressSettings, outCompressedTracks, stats);

	return result;
}

到这里模型的骨骼信息和动画信息都封装好了

渲染动画

动画相关组件

struct DynamicMeshComponent
{
	AssetHandle MeshSource = 0;
};
struct SubmeshComponent
{
	AssetHandle Mesh;
	std::vector<UUID> BoneEntityIds; 
	uint32_t SubmeshIndex = 0;
	bool Visible = true;

	SubmeshComponent() = default;
	SubmeshComponent(const SubmeshComponent& other)
		: Mesh(other.Mesh), BoneEntityIds(other.BoneEntityIds), SubmeshIndex(other.SubmeshIndex), Visible(other.Visible)
	{
	}
	SubmeshComponent(AssetHandle mesh, uint32_t submeshIndex = 0)
		: Mesh(mesh), SubmeshIndex(submeshIndex)
	{
	}
};
struct AnimationComponent
{
    AssetHandle Mesh;
    const Animation* CurrentAnimation; 
    float CurrentTime = 0.0f;
    bool IsLooping = true; 
    Pose CurrentPose;      

    std::vector<UUID> BoneEntityIds; 
    AnimationComponent() = default;
    AnimationComponent(AssetHandle mesh):Mesh(mesh)
    {
    }
};

收集动画Mesh

定义骨骼UBO，同实例化渲染的设计，通过MeshKey来区分不一样的DrawCall，所有Submesh只要材质一样就会被实例化调用，另外用StorageBufferSet平铺骨骼变换矩阵

		using BoneTransforms = std::array<glm::mat4, 100>; // Note: 100 == MAX_BONES from the shaders
		struct BoneTransformsMapData
		{
			std::vector<BoneTransforms> BoneTransformsData;
			uint32_t BoneTransformsBaseIndex = 0;
		};
		std::map<MeshKey, BoneTransformsMapData> m_MeshBoneTransformsMap;
		Ref<StorageBufferSet> m_SBSBoneTransforms;

Shader中，通过a_BoneIndices传递该顶点绑定的骨骼，a_BoneWeights传递权重。这些信息在Mesh导入时已经被封装。通过顶点缓冲区导入

#version 450 core
#ifdef VERTEX_SHADER
layout(binding = 0) uniform UniformBufferObject {
    mat4 view;
    mat4 proj;
	float width;
	float height;
	float Near;
	float Far;
} ubo;
const int MAX_BONES = 100;
const int MAX_ANIMATED_MESHES = 1024;
layout(location = 0) in vec3 inPosition;
layout(location = 1) in vec3 inNormal;
layout(location = 2) in vec3 inTangent;
layout(location = 3) in vec3 inBinormal; 
layout(location = 4) in vec2 inTexCoord;
layout(location = 5) in vec4 a_MRow0;
layout(location = 6) in vec4 a_MRow1;
layout(location = 7) in vec4 a_MRow2;
// Bone influences
layout(location = 8) in ivec4 a_BoneIndices;
layout(location = 9) in vec4 a_BoneWeights;
layout (std140, set = 0, binding = 1) readonly buffer BoneTransforms
{
	mat4 BoneTransforms[MAX_BONES * MAX_ANIMATED_MESHES];
} r_BoneTransforms;

layout(push_constant) uniform BoneTransformIndex
{
	uint Base;
} u_BoneTransformIndex;


layout(location = 0) out vec3 fragWorldPos;
layout(location = 1) out vec3 fragWorldNormal;
layout(location = 2) out vec3 fragWorldTangent;
layout(location = 3) out vec3 fragWorldBinormal;
layout(location = 4) out vec2 fragTexCoord;

void main() {
    mat4 transform = mat4(
		vec4(a_MRow0.x, a_MRow1.x, a_MRow2.x, 0.0),
		vec4(a_MRow0.y, a_MRow1.y, a_MRow2.y, 0.0),
		vec4(a_MRow0.z, a_MRow1.z, a_MRow2.z, 0.0),
		vec4(a_MRow0.w, a_MRow1.w, a_MRow2.w, 1.0)
	);
	mat4 boneTransform = r_BoneTransforms.BoneTransforms[(u_BoneTransformIndex.Base + gl_InstanceIndex) * MAX_BONES + a_BoneIndices[0]] * a_BoneWeights[0];
	boneTransform     += r_BoneTransforms.BoneTransforms[(u_BoneTransformIndex.Base + gl_InstanceIndex) * MAX_BONES + a_BoneIndices[1]] * a_BoneWeights[1];
	boneTransform     += r_BoneTransforms.BoneTransforms[(u_BoneTransformIndex.Base + gl_InstanceIndex) * MAX_BONES + a_BoneIndices[2]] * a_BoneWeights[2];
	boneTransform     += r_BoneTransforms.BoneTransforms[(u_BoneTransformIndex.Base + gl_InstanceIndex) * MAX_BONES + a_BoneIndices[3]] * a_BoneWeights[3];

    vec4 worldPos = transform * boneTransform * vec4(inPosition, 1.0);
    fragWorldPos = worldPos.xyz;
		
    mat3 modelRot = mat3(transform);
    mat3 boneRot = mat3(boneTransform);
    fragWorldNormal = normalize(modelRot * boneRot * inNormal);
    fragWorldTangent = normalize(modelRot * boneRot * inTangent);
    fragWorldBinormal = normalize(modelRot * boneRot * inBinormal);
    fragTexCoord = inTexCoord;
    gl_Position = ubo.proj * ubo.view * worldPos;
}
#endif