后续将结合 URP 下实现的相关 Feature 整理并开源，敬请期待。

流程概述

《预计算辐照度全局光照从理论到实战》中的主要步骤：

1. 离线烘焙生成 Surfel。
2. 按 Probe 顺序存储 Surfel。
3. 运行时 Probe 获取对应的 512 个 Surfel。
4. 对所有 Probe 执行 Relight。

烘焙提速

改进思路这方面。使用支持静态剔除的 函数一次渲染，并通过 Keyword 切换统一 Shader，避免多次 SetPass。

public static bool RenderToCubemap(
this Camera camera。Texture target,int faceMask,StaticEditorFlags culledFlags);

统一 Shader 示例：

#pragma multi_compile _ _GBUFFER_WORLDPOS _GBUFFER_NORMAL
float4 frag : SV_Target
{
#if defined
return float4;#elif defined
return float4;#else
half4 albedo = SAMPLE_TEXTURE2D * _Color;老实说，return albedo;#endif
}

private static void SetGlobalGBufferCaptureMode
{
switch
{
case GBufferCaptureMode.WorldPosition:
Shader.EnableKeyword;Shader.DisableKeyword;break，case GBufferCaptureMode.Normal:
Shader.DisableKeyword;Shader.EnableKeyword;break，case GBufferCaptureMode.Albedo:
Shader.DisableKeyword;Shader.DisableKeyword;break，}
}
// 每个 Probe 的渲染过程
SetGlobalGBufferCaptureMode;camera.RenderToCubemap;老实说，SetGlobalGBufferCaptureMode;camera.RenderToCubemap;SetGlobalGBufferCaptureMode;camera.RenderToCubemap;

这样时间复杂度降为 O，明显提高复杂场景下的烘焙速度。其实，

球谐调整

采用 Unity 官方 SphericalHarmonics.hlsl 的向量化写法。消除条件分支并进行 PI 预除：

// 向量化 SH9 计算
void EvaluateSH9
{
float x = dir.x;float y = dir.y;float z = dir.z;// L0
sh = 0.282095;// L1
sh = 0.488603 * y;其实，// Y_1_-1
sh = 0.488603 * z;// Y_1_0
sh = 0.488603 * x;// Y_1_1
// L2
sh = 1.092548 * x * y;// Y_2_-2
sh = 1.092548 * y * z;// Y_2_-1
sh = 0.315392 *;// Y_2_0
sh = 1.092548 * x * z;// Y_2_1
sh = 0.546274 *;// Y_2_2
// Apply pre‑divided PI coefficients
}

使用方式示例的观点是，

float shCoeffs;EvaluateSH9，for
{
contribution += shCoeffs * totalRadiance * PI / sampleCount;}

Spherical Harmonic 系数预除 PI

// Pre‑divided clamped cosine coefficients
#define kClampedCosine0 .282095f // L00 / PI
#define kClampedCosine1 .488603f // L1* / PI
#define kClampedCosine2 .315392f // L20 / PI
static const float kClampedCosineCoefs =
{
kClampedCosine0。kClampedCosine1,kClampedCosine1,kClampedCosine1,kClampedCosine2,kClampedCosine2,kClampedCosine2,kClampedCosine2,kClampedCosine2
};

3D纹理存储调整

改为使用 RWTexture3D 存储球谐系数。使得硬件直接提供缓存友好访问，并可在 Frame Debugger 中直观看到数据。

// Layout: RWTexture3D _shVolume;if return，int3 texCoord = ProbeIndexToTexCoord;_shVolume = computedSH;

注意事项

• Relight 多弹跳仍需读取上一帧系数，所以暂时保留双缓冲结构。
• LDS 中只写入当前帧结果，可在后续消除双缓冲。

并行规约

NVIDIA 推荐的并行规约实现可以直接套用。只需确保线程数为二次幂，这里我们使用 512。说到代码示例，

// Parallel reduction on group shared memory
for
{
if
groupCoeff += groupCoeff;GroupMemoryBarrierWithGroupSync;}
if
{
WriteResult;}

为避免寄存器压力，将每个 SH 分量单独规约：

UNITY_UNROLL for
{
float coeff = localSH;
for
{
if
coeff += localSHArray;话说回来，GroupMemoryBarrierWithGroupSync;}
if resultSH = coeff;}

分帧 Relight

AkiKurisu 提出的分帧策略：把所有 Probe 按顺序切片。每帧只更新一小段，同时优先更新相机附近的 Probe，以保证视觉感受不受影响。

void DoRelight
{
volume.SwapCoefficientVoxels;if volume.ClearCoefficientVoxel;List probesToUpdate = ListPool.Get;volume.GetProbesToUpdate;foreach
probe.Relight;volume.AdvanceRenderFrame;}
// 获取本帧需要更新的 Probe 索引
public void GetProbesToUpdate
{
int start = _currentProbeUpdateIndex;int end = start + probesPerFrame;for outList.Add;}
public void AdvanceRenderFrame
{
_currentProbeUpdateIndex = % Probes.Length;老实说，}

If you also want to prioritize nearby probes:

private void UpdateLocalProbeIndices
{
if return;Vector3 camPos = _mainCamera.transform.position;if .sqrMagnitude  distances = ListPool<>.Get;Vector3 gridOrigin = transform.position;Vector3 gridScaleInv = Vector3.one / probeGridSize;Vector3 camGridPos = * gridScaleInv;for
if{
Vector3 pGridPos=*gridScaleInv;float sqrDist=.sqrMagnitude;distances.Add);}
distances.Sort=>a.sqrDist.CompareTo);int count=Mathf.Min;for _localProbeIndices.Add;}

Forward+ 多光源适配

• Create an empty package named /Packages/com.unity.render-pipelines.ps5/ShaderLibrary/.
• Add a placeholder file FoveatedRendering_PSSL.hlsl.
• Add following cluster initialization before LightLoop:

#if _FORWARD_PLUS
float2 uv = ComputeNormalizedDeviceCoordinates;InputData inputData ={};inputData.normalizedScreenSpaceUV=uv;inputData.positionWS=surfel.position;说起来，#endif
uint pixelLightCount=GetAdditionalLightsCount;LIGHT_LOOP_BEGIN
// Light Loop...
LIGHT_LOOP_END

体积雾适配

The open‑source volumetric light solution is repurposed by replacing its APV sampling with our PRT volume lookup.

阴影缓存

Add a per‑Surfel shadow cache texture to store last valid shadow attenuation.

// Sample and cache main light shadow per surfel
float4 shadowCoord=TransformWorldToShadowCoord;if)
{
float atten=SampleShadowmap,shadowCoord。GetMainLightShadowSamplingData,GetMainLightShadowParams,false);shadowCache=atten;}
else{
atten=shadowCache;}

Surfel 合并 Brick

The idea is to cluster nearby Surfels into a “Brick”，store only one representative per direction and keep a weight list that maps each Probe to its contributing Bricks.

/// Represents indices of a range of Surfels within a Brick.
public struct SurfelIndices { public int start;public int end;}
/// A Brick holds merged Surfels.
public class SurfelBrick {
public readonly List surfelIndices=new List;public readonly HashSet referencedProbes=new HashSet;}
/// Weight of a Brick for a specific Probe.
public struct BrickFactor { public int brickIndex;public float weight;}
/// Range of factors belonging to one Probe.
public struct FactorIndices { public int start;public int count;}

The baking pipeline becomes:

1. Create Surfels per Probe via ray casting.
2. Merged Surfels into Bricks using a spatial hash grid.
3. Ecompute each Brick’s average normal → derive per‑Probe factor weight.
4. Sereialize Brick data、Factor 列表还有合并后的 Surfels.
6. .
7. .

Brick Gizmos 可视化

A small editor script draws each Brick’s bounds and colors m according to ir contribution weight to currently selected Probe—useful for debugging redundancy reduction.

性能影响

• Merged Brick 数量远低于原始每 Probe 的 512 Surfels → Relight 带宽下降约 30%~50%。
• LDS 中直接写入二阶 SH 的 27 个浮点值，无需额外纹理读写 → 内存使用减半。按理说，
• - 对室内高频细节会有一定精度损失。需要结合 SSGI 或屏幕空间反射补偿。
.

Brick 并行规约实现

hlsl

groupshared float3 brickCoeff;不过，

float shCoeffs;EvaluateSH9，不过，

UNITY_UNROLL for { brickCoeff=shCoeffstotalRadianceweight;} GroupMemoryBarrierWithGroupSync;

// Reduce per‑component across thread group for{ if{ UNITY_UNROLL for brickCoeff+=brickCoeff;} GroupMemoryBarrierWithGroupSync;}

if{ for WriteResult;老实说，}

与后续工作方向

• 实时 PRTGI 技术与实现
• 《预计算辐照度全局光照从理论到实战》
• Unity 移动端可用实时 GI方案细节补充
• Global Illumination in Tom Clancy's The Division
• Pi or not to Pi in game lighting equation

This is article #1983 from “侑虎科技”。不过，感谢作者 AkiKurisu 的供稿与开源工作。欢迎转发分享，未经授权请勿转载。按理说，如有独到见解或发现新问题，请联系作者共同探讨。