Stochastic Screen Space Reflections(三):实现与优化

短命女 2022-05-06 04:00 279阅读 0赞

必须在延迟管线下渲染，且在不透明后，半透明前渲染

本文在[Xerxes1138][]的基础上进行改进

# 实现 #

Xerxes1138的实现比Siggraph2015的步骤省略了不少内容

![70][]

少了Tile分类和Hi-Z Trace

首先逐个pass来看他的实现过程

## Recursive Pass ##

//uv-速度(回到上一帧uv) 
    float2 prevUV = uv - velocity; 
    //上一帧渲染结果 
    float4 sceneColor = SAMPLE_TEXTURE2D(_MainTex, sampler_MainTex,prevUV); 
    return sceneColor;

返回上一帧着色结果，包括上一帧的StoSSR结果，这样就能模拟出多Bounces的结果，接下来的RayTrace采样就是用这张RT

## RayCast Pass ##

该Pass返回RayTrace hit到的点的位置为一张RT，RayMask为另一张RT表示没有反射到物体的情况

此处用的RayTrace类型只有一种，就是RayMarch获取hit点，这个方法精度是不如Hi-Z的，原文的步骤有一个块的分类，对重要的部分采用Hi-Z采样

![70 1][]

### Hi-Z ###

关于hierarchical tracing，在GPU Pro5中有详细介绍，

需要生成一个visible buffer的Mip Chain(\[Hermanns15\])

![70 2][]

![70 3][]

在Sig15里面的介绍：

![70 4][]

![70 5][]

![70 6][]

![70 7][]

![70 8][]

![70 9][]

![70 10][]

整体来讲就是一个先加速后减速的trace过程，这个过程减少了每步都是一小步进行检测的消耗，结果也比较准确基本和每步一小步结果一样

# MipMapBlur Pass #

这一步就是构造一个Mip Chain，来模拟不同粗糙度的反射结果，使用5个Mip级别与Surge中相同

![70 11][]

Xerxes1138中横向Blur一次+纵向Blur一次算一个Mip，相当于10次Pass，此处可以直接用mipmap处理比较省

# Resolve Pass #

此步骤使用Trace返回的hit点的位置与Mip chain根据粗糙度进行最终颜色处理，生成反射颜色

如果开启ReUse，那么就用相邻点(Xerxes1138中是4个)混合，来模拟Cone Trace

![70 12][]

# Temporal Pass #

该步骤做TAA处理，可以大量降低噪声

原理就是与前一帧渲染结果(CPU传入，不是UV算的)做混合，前一阵渲染结果需要过滤，需要clamp在周围8个采样点包括当前点一共9个点的明暗范围之内

//
    //pass:temporal(第五步)
    //返回:TAA后结果
    //与前一帧渲染结果做混合，前一阵渲染结果需要过滤，需要clamp在周围8个采样点包括当前点一共9个点的明暗范围之内
    //
    	void temporal (VaryingsDefault i, out half4 reflection : SV_Target)
    	{	
    		float2 uv = i.texcoordStereo;
    		//获取信息
    		//深度/屏幕空间位置/速度/_RayCast？
    		float depth = GetDepth(/*_CameraDepthTexture,*/ uv);
    		float3 screenPos = GetScreenPos(uv, depth);
    		float2 velocity = GetVelocity(uv); // 5.4 motion vector
    		float2 hitPacked = SAMPLE_TEXTURE2D_LOD(_RayCast, sampler_RayCast, uv,0);//  tex2Dlod(_RayCast, float4(uv, 0.0, 0.0));
    
    		float2 averageDepthMin = min(hitPacked, velocity);
    		float2 averageDepthMax = max(hitPacked, velocity);
    
    		//velocity = clamp(velocity, averageDepthMin, averageDepthMax);
    		//获取上一帧UV
    		float2 prevUV = uv - velocity;
    		//当前RT
    		float4 current =SAMPLE_TEXTURE2D(_MainTex, sampler_MainTex,uv);
    		//上一帧RT，用上一帧UV取
    		float4 previous = SAMPLE_TEXTURE2D(_PreviousBuffer, sampler_PreviousBuffer,prevUV);
    		//一个像素长宽
    		float2 du = float2(1.0 / _ScreenSize.x, 0.0);
    		float2 dv = float2(0.0, 1.0 / _ScreenSize.y);
    		//一个点周围8个采样
    		float4 currentTopLeft =SAMPLE_TEXTURE2D(_MainTex, sampler_MainTex,uv.xy - dv - du);
    		float4 currentTopCenter =SAMPLE_TEXTURE2D(_MainTex, sampler_MainTex, uv.xy - dv);
    		float4 currentTopRight = SAMPLE_TEXTURE2D(_MainTex, sampler_MainTex,uv.xy - dv + du);
    		float4 currentMiddleLeft = SAMPLE_TEXTURE2D(_MainTex, sampler_MainTex,uv.xy - du);
    		float4 currentMiddleCenter = SAMPLE_TEXTURE2D(_MainTex, sampler_MainTex,uv.xy);
    		float4 currentMiddleRight = SAMPLE_TEXTURE2D(_MainTex, sampler_MainTex, uv.xy + du);
    		float4 currentBottomLeft = SAMPLE_TEXTURE2D(_MainTex, sampler_MainTex,uv.xy + dv - du);
    		float4 currentBottomCenter = SAMPLE_TEXTURE2D(_MainTex, sampler_MainTex,uv.xy + dv);
    		float4 currentBottomRight = SAMPLE_TEXTURE2D(_MainTex, sampler_MainTex,uv.xy + dv + du);
    		//取最暗采样点
    		float4 currentMin = min(currentTopLeft, min(currentTopCenter, min(currentTopRight, min(currentMiddleLeft, min(currentMiddleCenter, min(currentMiddleRight, min(currentBottomLeft, min(currentBottomCenter, currentBottomRight))))))));
    		//取最明采样点
    		float4 currentMax = max(currentTopLeft, max(currentTopCenter, max(currentTopRight, max(currentMiddleLeft, max(currentMiddleCenter, max(currentMiddleRight, max(currentBottomLeft, max(currentBottomCenter, currentBottomRight))))))));
    
    		float scale = _TScale;
    		//平均混合最明和最暗
    		float4 center = (currentMin + currentMax) * 0.5f;
    		//缩放明暗差
    		currentMin = (currentMin - center) * scale + center;
    		currentMax = (currentMax - center) * scale + center;
    		//将上一帧Buffer clamp到 当前采样点 min max范围内
    		previous = clamp(previous, currentMin, currentMax);
    		//混合当前RT和移动前RT颜色，/velocity移动距离(速度)越大，约接近当前RT值/velocity移动距离(速度)越小，约接近移动前RT值
        	reflection = lerp(current, previous, saturate(_TResponse *  (1 - length(velocity) * 8)) );
    	}

# Combine Pass #

混合场景颜色和反射颜色

根据mask混合\_CameraReflectionsTexture和反射计算结果的值 再加上 (场景颜色-\_CameraReflectionsTexture)

float4 sceneColor = SAMPLE_TEXTURE2D(_MainTex, sampler_MainTex,uv);//tex2D(_MainTex,  uv);
    //场景色-cubeMap颜色
    sceneColor.rgb = max(1e-5, sceneColor.rgb - cubemap.rgb);
    //最终结果，混合反射结果和cubemap值
    sceneColor.rgb += lerp(cubemap.rgb, reflection.rgb, mask);

此处的场景颜色为当前帧渲染source

# 问题 #

## 射线ReUse产生Artifact ##

![70 13][]

![70 14][]

此处可以加个判断，如果邻居点与当前点法线角度差太多就不混合结果。因为如果是粗糙表面，cone trace也有一个角度范围限制的，模拟一个面的反射，不会角度相差很多。

![70 15][]

![70 16][]

但是如果相邻点在角度差别不大的两个物体，依然会有artifact产生

![70 17][]

参考：

1.\[Hermanns15\] Lukas Hermanns "Screen space cone tracing for glossy reflections"  
http://publica.fraunhofer.de/documents/N-336466.html

\----- by wolf96 [https://blog.csdn.net/wolf96][https_blog.csdn.net_wolf96]

[Xerxes1138]: https://github.com/Xerxes1138/StochasticScreenSpaceReflection
[70]: /images/20220506/2ae74488f39c42639ccca3cae57c38f4.png
[70 1]: /images/20220506/ca78f9d41e374dd79c9b6c8787591d6d.png
[70 2]: /images/20220506/1b76d60d5a9f4168bc84d5288549c6c9.png
[70 3]: /images/20220506/2e4ca8b02c0943deb5c06be641baf731.png
[70 4]: /images/20220506/724033e549044853b209aea60535d84a.png
[70 5]: /images/20220506/611dc9d331884c16b9beb4ca8502a56d.png
[70 6]: /images/20220506/d2c8890ba5b0474e95acabb0d0da7744.png
[70 7]: /images/20220506/7e35a84de55d4973b591ade4cdc5b315.png
[70 8]: /images/20220506/b8ba694073104ff8bc642a2063d0bba3.png
[70 9]: /images/20220506/eb70d88804b6426188a0d5f13b1d227d.png
[70 10]: /images/20220506/8cdf7e12aedf476a9bef22ad27480e64.png
[70 11]: /images/20220506/4996a5c65934452bb9f47c1aebd5e8ea.png
[70 12]: /images/20220506/42b427a152c6423a8dc24cdd6607f71a.png
[70 13]: /images/20220506/9b9e1c630eef43558bb7729774271e4a.png
[70 14]: /images/20220506/c7ae699542db4432a96b35300cd730f4.png
[70 15]: /images/20220506/62eacca2bcaf432e87545a7c3e01786a.png
[70 16]: /images/20220506/df5a6dd3b33547868ed1aa49a5a0face.png
[70 17]: /images/20220506/eb9c0df6ee0f4aa68b1a4ae86a66810b.png
[https_blog.csdn.net_wolf96]: https://blog.csdn.net/wolf96