|
|
(Github正常排版: 写写简单的HBAO)
<hr/>0. 起因
URP有SSAO, HDRP有GTAO. 所以摆烂学一个HBAO. 官方的Github地址
下面是效果图.
下面是URP的SSAO. 可以发现HBAO会软很多, 慢慢淡出.
下面是自己写的GTAO, 调参没有调对. GTAO因为多了多次弹射的结果, 感觉更黑了.
下面随便写的RayTracingAO, 也没有仔细的调参. 但是视觉效基本上和上面的差不多.
自己调参可能有点不准确, 这是官方给的SSAO和HBAO的对比图.
HBAO对比SSAO采样次数更少, 效果也好很多, 不过使用了多次三角函数. 虽然SSAO可以变成TSSAO来减少采样和降低噪点.
同时我在找资料的时候还发现CACAO, 算是SSAO的升级, 效果也不错.Github地址
<hr/>1. 原理
HBAO, Image-Space Horizon-Based Ambient Occlusion, 水平基准环境光遮蔽, 是一项英伟达于2008年提出的SSAO衍生版, 效果比SSAO更好. 文章地址
YiQiuuu的有篇文章是关于HBAO原理和实现, 讲的详细且不错, 文章地址. 这里直接快速引用概括一下.
- 屏幕上的每一个像素, 做一个四等分的四条射线, 然后随机旋转一下. 每一个像素的随机角度不能一样, 否则效果很怪/是错的. 这里的四等分也可以是六等分, 八等分...
2. 对于任意一条射线, 沿着射线方向生成一个一维的高度. 然后RayMarch根据深度做找到一个最大的水平角(Horizon Angle).
3. 根据点P和它的法线(面法线), 计算它的切面角(Tangent Angle).
4. 根据Horizon Angle和Tangent Angle, 得到AO. AO = sin(h) - sin(t).
至于为什么AO=角度值的差?
我的理解(个人理解)是 周围的东西对比当前点越高, 则表示光被周围东西遮挡的越多, 则越暗.
为什么用面法线, 而不是顶点插值法线?
看官网的PPT是说 如果用顶点插值法线去计算会得到错误的遮挡.
如果我们用的是顶点插值法线, 当P在拐弯位置的时候, 顶点插值法线和面法线不一致. 计算的半球起始位置就可能会是错的.
而根据View Space 利用ddx/ddy重新生成的法线是对的. 之前的SSAO篇中有介绍怎么重新生成法线.
但是我下面的代码还是会用NormalRT, 首先为了提高性能.
下面是我用ddx/ddy重新生成的法线出的效果. 发现会出现奇怪的死黑的边缘和锯齿.
如果用的是比较复杂算法生成NormalRT, 看着效果和Gbuffer的NormalRT产生的效果也差不多.
而且物体一般都有NormalMap, 用深度图和算法重新生成的NormalRT是没有NormalMap的. 比如地表的石子用的是NormalMap, 就会出现下面的效果.
因此我这里还是用的GBuffer的Normal 或者 DepthNormals Pass 生成的NormalRT.
其实正确的应该是用BentNormalRT. BentNormal大体指光线大概率通过的平均方向/不被其他物体遮挡的方向. 因此生成这个也更麻烦, 业界做法常常是离线烘焙好贴图.
<hr/>2. C#
个人习惯, 先写C#吧.
2.1 RenderSettings
创建一个C#文件HBAORenderFeature.cs. 先写RenderSettings.
using UnityEngine;
using UnityEngine.Rendering;
using UnityEngine.Rendering.Universal;
[System.Serializable]
public class HBAORenderSettings
{
[Range(0.0f, 1.0f)] public float intensity = 1.0f;
[Range(0.25f, 5.0f)] public float radius = 1.2f;
[Range(16f, 256f)] public float maxRadiusPixels = 256;
[Range(0.0f, 0.5f)] public float angleBias = 0.05f;
[Min(0)] public float maxDistance = 150.0f;
[Min(0)] public float distanceFalloff = 50.0f;
[Range(0.0f,16.0f)] public float sharpness = 8.0f;
}
intensity: AO强度
radius: 射线半径
maxRadiusPixels: 最大半径的像素数量
angleBias: 角度阈值
maxDistance: AO的最大距离, 超过这个距离就没有AO了
distanceFalloff: AO的距离衰减, 淡出用
sharpness: 模糊深度权重
2.2 RenderFeature
然后继续在C#文件HBAORenderFeature.cs中, 写RenderFeature. HBAORenderPass 在后面补充.
因为URP的生命周期越来越神奇 存在反复调用, 所以我用 OnCreate() 和 flag 来管理.
这里只是demo, renderPassEvent我是随便写的. 比如这里是RenderPassEvent.BeforeRenderingPostProcessing.
using UnityEngine;
using UnityEngine.Rendering;
using UnityEngine.Rendering.Universal;
[System.Serializable]
public class HBAORenderSettings
{
...
}
public class HBAORenderFeature : ScriptableRendererFeature
{
public Shader effectShader;
public HBAORenderSettings renderSettings;
private bool needCreate;
private HBAORenderPass renderPass;
private Material effectMat;
public override void Create()
{
needCreate = true;
}
protected override void Dispose(bool disposing)
{
CoreUtils.Destroy(effectMat);
if (renderPass != null)
{
renderPass.OnDestroy();
renderPass = null;
}
}
public void OnCreate()
{
if (!needCreate)
{
return;
}
needCreate = false;
if (renderPass == null)
{
renderPass = new HBAORenderPass()
{
renderPassEvent = RenderPassEvent.BeforeRenderingPostProcessing,
};
}
if (effectMat == null || effectMat.shader != effectShader)
{
CoreUtils.Destroy(effectMat);
if (effectShader != null)
{
effectMat = CoreUtils.CreateEngineMaterial(effectShader);
}
}
renderPass.OnInit(effectMat, renderSettings);
}
public override void AddRenderPasses(ScriptableRenderer renderer, ref RenderingData renderingData)
{
if (effectShader == null)
{
return;
}
OnCreate();
renderer.EnqueuePass(renderPass);
}
}
到这里RenderFeature基本就写完了.
2.3 RenderPass
创建HBAORenderPass.cs文件. 先写一个基础框架.
using UnityEngine;
using UnityEngine.Rendering;
using UnityEngine.Rendering.Universal;
public class HBAORenderPass : ScriptableRenderPass
{
private const string k_tag = &#34;HBAO&#34;;
private HBAORenderSettings settings;
private Material effectMat;
public HBAORenderPass()
{
profilingSampler = new ProfilingSampler(k_tag);
}
public void OnInit(Material _effectMat, HBAORenderSettings _renderSettings)
{
effectMat = _effectMat;
settings = _renderSettings;
}
public void OnDestroy()
{
}
public override void Configure(CommandBuffer cmd, RenderTextureDescriptor cameraTextureDescriptor)
{
}
public override void Execute(ScriptableRenderContext context, ref RenderingData renderingData)
{
var cmd = CommandBufferPool.Get();
using (new ProfilingScope(cmd, profilingSampler))
{
}
context.ExecuteCommandBuffer(cmd);
CommandBufferPool.Release(cmd);
}
}
从前面知道我们需要depthRT和normalRT, 所以在Configure中配置ConfigureInput.
public class HBAORenderPass : ScriptableRenderPass
{
...
public override void Configure(CommandBuffer cmd, RenderTextureDescriptor cameraTextureDescriptor)
{
ConfigureInput(ScriptableRenderPassInput.Depth | ScriptableRenderPassInput.Normal);
}
...
}
然后我们需要一个随机旋和射线长度的噪音图, 这里我偷懒用了ComputeBuffer. 但是建议离线把Texture保存下来传入.
创建一个变量 ComputeBuffer noiseCB , 和 GenerateNoise 方法.
noise.x: 随机初始角度
noise.y: 射线随机初始化长度
public class HBAORenderPass : ScriptableRenderPass
{
...
private ComputeBuffer noiseCB;
private HBAORenderSettings settings;
private Material effectMat;
...
public void OnInit(Material _effectMat, HBAORenderSettings _renderSettings)
{
effectMat = _effectMat;
settings = _renderSettings;
if (noiseCB != null)
{
noiseCB.Release();
}
Vector2[] noiseData = GenerateNoise();
noiseCB = new ComputeBuffer(noiseData.Length, sizeof(float) * 2);
noiseCB.SetData(noiseData);
}
public void OnDestroy()
{
if (noiseCB != null)
{
noiseCB.Release();
noiseCB = null;
}
}
...
public override void Execute(ScriptableRenderContext context, ref RenderingData renderingData)
{
...
}
private Vector2[] GenerateNoise()
{
Vector2[] noises = new Vector2[4 * 4];
for (int i = 0; i < noises.Length; i++)
{
float x = Random.value;
float y = Random.value;
noises = new Vector2(x, y);
}
return noises;
}
}
然后就是Shader属性ID. 这里直接全部一把梭写完了, 不用来回折腾, 也方便自动补全.
public class HBARenderPass : ScriptableRenderPass
{
private const string k_tag = &#34;HBAO&#34;;
private static readonly int hbaoRT_ID = Shader.PropertyToID(&#34;_HBAORT&#34;);
private static readonly int hbaoBlurRT_ID = Shader.PropertyToID(&#34;_HBAOBlurRT&#34;);
private static readonly int noiseCB_ID = Shader.PropertyToID(&#34;_NoiseCB&#34;);
private static readonly int aoTex_ID = Shader.PropertyToID(&#34;_AOTex&#34;);
private static readonly int intensity_ID = Shader.PropertyToID(&#34;_Intensity&#34;);
private static readonly int radius_ID = Shader.PropertyToID(&#34;_Radius&#34;);
private static readonly int negInvRadius2_ID = Shader.PropertyToID(&#34;_NegInvRadius2&#34;);
private static readonly int maxRadiusPixels_ID = Shader.PropertyToID(&#34;_MaxRadiusPixels&#34;);
private static readonly int angleBias_ID = Shader.PropertyToID(&#34;_AngleBias&#34;);
private static readonly int aoMultiplier_ID = Shader.PropertyToID(&#34;_AOMultiplier&#34;);
private static readonly int maxDistance_ID = Shader.PropertyToID(&#34;_MaxDistance&#34;);
private static readonly int distanceFalloff_ID = Shader.PropertyToID(&#34;_DistanceFalloff&#34;);
private static readonly int blurSharpness_ID = Shader.PropertyToID(&#34;_BlurSharpness&#34;);
private static readonly int blurDeltaUV_ID = Shader.PropertyToID(&#34;_BlurDeltaUV&#34;);
...
}
在 Execute(ScriptableRenderContext context, ref RenderingData renderingData) 方法中, 把Settings参数和NoiseCB等传输给GPU.
radius, 跟屏幕尺寸比例有关.
maxRadiusPixels, 跟屏幕分辨率有关.
aoMultiplier, 跟bias有关系, 因为bias会减弱ao, 所以这个做补偿.
public override void Execute(ScriptableRenderContext context, ref RenderingData renderingData)
{
var cmd = CommandBufferPool.Get();
using (new ProfilingScope(cmd, profilingSampler))
{
int width = renderingData.cameraData.cameraTargetDescriptor.width;
int height = renderingData.cameraData.cameraTargetDescriptor.height;
float fov = renderingData.cameraData.camera.fieldOfView;
float tanHalfFovY = Mathf.Tan(0.5f * fov * Mathf.Deg2Rad);
cmd.SetGlobalBuffer(noiseCB_ID, noiseCB);
cmd.SetGlobalFloat(intensity_ID, settings.intensity);
cmd.SetGlobalFloat(radius_ID, settings.radius * 0.5f * height / (2.0f * tanHalfFovY));
cmd.SetGlobalFloat(negInvRadius2_ID, -1.0f / (settings.radius * settings.radius));
float maxRadiusPixels = settings.maxRadiusPixels * Mathf.Sqrt((width * height) / (1080.0f * 1920.0f));
cmd.SetGlobalFloat(maxRadiusPixels_ID, Mathf.Max(16, maxRadiusPixels));
cmd.SetGlobalFloat(angleBias_ID, settings.angleBias);
cmd.SetGlobalFloat(aoMultiplier_ID, 2.0f / (1.0f - settings.angleBias));
cmd.SetGlobalFloat(maxDistance_ID, settings.maxDistance);
cmd.SetGlobalFloat(distanceFalloff_ID, settings.distanceFalloff);
}
context.ExecuteCommandBuffer(cmd);
CommandBufferPool.Release(cmd);
}
然后就是申请RT, SetRT, 绘制HBAO, 别忘了释放RT.
Blur和Combine的模块之后再补充.
public override void Execute(ScriptableRenderContext context, ref RenderingData renderingData)
{
...
using (new ProfilingScope(cmd, profilingSampler))
{
...
cmd.SetGlobalFloat(distanceFalloff_ID, settings.distanceFalloff);
cmd.GetTemporaryRT(hbaoRT_ID, width, height, 0, FilterMode.Bilinear, RenderTextureFormat.R8, RenderTextureReadWrite.Linear);
cmd.SetRenderTarget(hbaoRT_ID);
CoreUtils.DrawFullScreen(cmd, effectMat, null, 0);
//TODO:Blur
//TODO:Combine
cmd.ReleaseTemporaryRT(hbaoRT_ID);
}
...
}
这样简单的HBAO C#基本写完了.
<hr/>3. AOShader
3.1 基础的框架
创建一个Shader文件 HBAO.shader. 写个空的框架进去.
Shader &#34;HBAO&#34;
{
SubShader
{
Tags
{
&#34;RenderType&#34;=&#34;Opaque&#34;
}
LOD 100
Cull Off
ZWrite Off
ZTest Always
Pass
{
Name &#34;HBAO&#34;
HLSLPROGRAM
#pragma vertex vert
#pragma fragment frag
#include &#34;Packages/com.unity.render-pipelines.core/ShaderLibrary/Common.hlsl&#34;
#include &#34;Packages/com.unity.render-pipelines.universal/ShaderLibrary/Core.hlsl&#34;
struct a2v
{
uint vertexID :SV_VertexID;
};
struct v2f
{
float4 pos:SV_Position;
float2 uv:TEXCOORD0;
};
v2f vert(a2v IN)
{
v2f o;
o.pos = GetFullScreenTriangleVertexPosition(IN.vertexID);
o.uv = GetFullScreenTriangleTexCoord(IN.vertexID);
return o;
}
half frag(v2f IN) : SV_Target
{
float2 uv = IN.uv;
return 1;
}
ENDHLSL
}
}
}
把C#的传入的参数写上.
还有我们需要DephtRT和NormalRT, 把它们的include也写上 DeclareDepthTexture.hlsl 和 DeclareNormalsTexture.hlsl.
同时定义方向等分个数(DIRECTIONS)和射线步进次数(STEPS). 这里直接大力出奇迹! 8等分, 6步进.
我这里偷懒用了Buffer!!!
Pass
{
Name &#34;HBAO&#34;
HLSLPROGRAM
...
#include &#34;Packages/com.unity.render-pipelines.universal/ShaderLibrary/Core.hlsl&#34;
#include &#34;Packages/com.unity.render-pipelines.universal/ShaderLibrary/DeclareDepthTexture.hlsl&#34;
#include &#34;Packages/com.unity.render-pipelines.universal/ShaderLibrary/DeclareNormalsTexture.hlsl&#34;
...
struct v2f
{
...
};
#define DIRECTIONS 8
#define STEPS 6
StructuredBuffer<float2> _NoiseCB;
float _Intensity;
float _Radius;
float _NegInvRadius2;
float _MaxRadiusPixels;
float _AngleBias;
float _AOMultiplier;
float _MaxDistance;
float _DistanceFalloff;
v2f vert(a2v IN)
{
...
}
...
}
3.2 数据准备
在循环判断HBAO之前还要在Fragment中做点数据准备.
先写一个 GetViewPos . 输入uv, uv采样深度图获得depth. uv和depth和UNITY_MATRIX_I_P 反算出viewPos. viewPos的z是near~far而不是0~1.
再写一个 FetchViewNormals . 采样之前ConfigInput的WorldNormal, 再转换到ViewSpace下. 注意反转YZ.
...
float _DistanceFalloff;
float3 GetViewPos(float2 uv)
{
float depth = SampleSceneDepth(uv);
float2 newUV = float2(uv.x, uv.y);
newUV = newUV * 2 - 1;
float4 viewPos = mul(UNITY_MATRIX_I_P, float4(newUV, depth, 1));
viewPos /= viewPos.w;
viewPos.z = -viewPos.z;
return viewPos.xyz;
}
float3 FetchViewNormals(float2 uv)
{
float3 N = SampleSceneNormals(uv);
N = TransformWorldToViewDir(N, true);
N.y = -N.y;
N.z = -N.z;
return N;
}
v2f vert(a2v IN)
{
...
}
回到 Frag .
得到当前像素的viewPos, 如果z>=_MaxDistance, 则无AO(return 1).
再得到当前像素的ViewNormal.
随机值. 根据屏幕坐标划分用NoiseCB得到noise. 因为这里是4x4=16的buffer, 所以是%4.
stepSize和stepAng. 步进的半径由深度决定, 近的时候放大, 远的时候缩小.
half frag(v2f IN) : SV_Target
{
float2 uv = IN.uv;
float3 viewPos = GetViewPos(uv);
if (viewPos.z >= _MaxDistance)
{
return 1;
}
float3 nor = FetchViewNormals(uv);
int noiseX = (uv.x * _ScreenSize.x - 0.5) % 4;
int noiseY = (uv.y * _ScreenSize.y - 0.5) % 4;
int noiseIndex = 4 * noiseY + noiseX;
float2 rand = _NoiseCB[noiseIndex];
float stepSize = min(_Radius / viewPos.z, _MaxRadiusPixels) / (STEPS + 1.0);
float stepAng = TWO_PI / DIRECTIONS;
...
}
3.3 循环迭代
下面就是循环迭代计算AO了. 先写一个二重循环, 循环角度和步进.
角度和射线步进长度 都需要 随机值 做起始. 而且射线的起始像素不包含自己本身的像素, 所以给个最小值1.
为什么要随机值?
因为效果看起来不会那么规则, 而且Blur之后的效果会更好. 可以看下面的两张图, 图一:无随机值且开启Blur 和 图二:有随机值开启Blur.
half frag(v2f IN) : SV_Target
{
...
float stepAng = TWO_PI / DIRECTIONS;
float ao = 0;
UNITY_UNROLL
for (int d = 0; d < DIRECTIONS; ++d)
{
float angle = stepAng * (float(d) + rand.x);
float cosAng, sinAng;
sincos(angle, sinAng, cosAng);
float2 direction = float2(cosAng, sinAng);
float rayPixels = frac(rand.y) * stepSize + 1.0;
UNITY_UNROLL
for (int s = 0; s < STEPS; ++s)
{
//TODO:ComputeAO
}
}
return ao;
}
在循环里面累加AO.
偏移UV, 得到新的偏移ViewPos.
利用 当前ViewPos, 当前Normal, 偏移ViewPos, 调用ComputeAO方法得到AO值, 方法在后面补充.
half frag(v2f IN) : SV_Target
{
...
UNITY_UNROLL
for (int s = 0; s < STEPS; ++s)
{
float2 snappedUV = round(rayPixels * direction) * _ScreenSize.zw + uv;
float3 tempViewPos = GetViewPos(snappedUV);
rayPixels += stepSize;
float tempAO = ComputeAO(viewPos, nor, tempViewPos);
ao += tempAO;
}
...
}
3.4 ComputeAO
完善ComputeAO方法.
这里的公式跟上面的PPT不一样, 但是思想大致相同.
首先上面PPT用的是重新生成的面法线, 所以要做起始角度的补偿. 但是这里用的是顶点插值生成的法线. 所以去掉了角度补偿.
接着我这里用的是累加每个Marching的AO值, 而文章是找到最大值角度值去计算AO.
NoV越大, 说明两个向量的角度越小, 即V靠近N, 周围的物体比较高, AO则越强. NoV 别忘了减去 settings 传入的 _AngleBias .
rsqrt(VoV) = 1 / sqrt(VoV) = 1 / sqrt(dot(VoV)) = 1 / length(V);
NoV = dot(n, v) * rsqrt(VoV) = dot(n, v / length(v)) = dot(n, normalize(v));
然后还要考虑到距离的衰减, 这里用了一个简单的公式: 1 - (dist^2 / maxDist^2).
为什么要 _AngleBias ?
不然在弧形区域会出现不连续的问题, 即出现一段一段的AO. 不过我自己试验了一下好像没有出现.
为什么要距离衰减?
距离衰减可以解决距离过渡导致的跳变的问题, 使其淡出更柔和. 而且还能解决在距离过大的时候, 但是还会计算AO的问题.
float Falloff(float distanceSquare)
{
return distanceSquare * _NegInvRadius2 + 1.0;
}
float ComputeAO(float3 p, float3 n, float3 s)
{
float3 v = s - p;
float VoV = dot(v, v);
float NoV = dot(n, v) * rsqrt(VoV);
return saturate(NoV - _AngleBias) * saturate(Falloff(VoV));
}
v2f vert(a2v IN)
{
...
}
...
3.5 强度
最后就是把上面累加的AO除以Count, 再把Settings的强度乘进去.
Settings的强度里面有考虑 AngleBias . 不然bias过大就会让AO显得很弱 偏白. 所以 引入 _AOMultiplier , AngleBias 越大, _AOMultiplier 会越小, 让AO越黑.
因为上面z>_MaxDistance 直接 return 1, 太过于突兀了. 所以这里再考虑一个AO距离的衰减.
half frag(v2f IN) : SV_Target
{
...
float ao = 0;
UNITY_UNROLL
for (int d = 0; d < DIRECTIONS; ++d)
{
...
}
//apply bias multiplier
ao *= _Intensity * _AOMultiplier / (STEPS * DIRECTIONS);
float distFactor = saturate((viewPos.z - (_MaxDistance - _DistanceFalloff)) / _DistanceFalloff);
ao = lerp(saturate(1 - ao), 1, distFactor);
return ao;
}
HBAO第一个pass写完基本就是下图这样, 充满噪点, 放大看就是一个格子一个格子的跳变. 后面就是横竖两次Blur就够了.
<hr/>4. Blur
4.1 C#
在HBAORenderPass.cs文件中, 修改Execute方法 添加blur pass.
再申请一个BlurRT. 把AORT作为Input, BlurRT作为Target, 进行一次Horizontal Blur. 再把BlurRT作为Input, AORT作为Target, 进行一次Vertical Blur.
最后别忘了销毁申请的BlurRT.
public class HBAORenderPass : ScriptableRenderPass
{
...
public override void Execute(ScriptableRenderContext context, ref RenderingData renderingData)
{
var cmd = CommandBufferPool.Get();
using (new ProfilingScope(cmd, profilingSampler))
{
...
CoreUtils.DrawFullScreen(cmd, effectMat, null, 0);
cmd.GetTemporaryRT(hbaoBlurRT_ID, width, height, 0, FilterMode.Bilinear, RenderTextureFormat.R8, RenderTextureReadWrite.Linear);
cmd.SetGlobalFloat(blurSharpness_ID, settings.sharpness);
cmd.SetGlobalVector(blurDeltaUV_ID, new Vector4(1.0f / width, 0, 0, 0));
cmd.SetGlobalTexture(aoTex_ID, hbaoRT_ID);
cmd.SetRenderTarget(hbaoBlurRT_ID);
CoreUtils.DrawFullScreen(cmd, effectMat, null, 1);
cmd.SetGlobalVector(blurDeltaUV_ID, new Vector4(0, 1.0f / height, 0, 0));
cmd.SetGlobalTexture(aoTex_ID, hbaoBlurRT_ID);
cmd.SetRenderTarget(hbaoRT_ID);
CoreUtils.DrawFullScreen(cmd, effectMat, null, 1);
//TODO:Combine
cmd.ReleaseTemporaryRT(hbaoRT_ID);
cmd.ReleaseTemporaryRT(hbaoBlurRT_ID);
}
context.ExecuteCommandBuffer(cmd);
CommandBufferPool.Release(cmd);
}
}
4.2 Shader框架
返回 HBAO.shader , 添加一个新的Blur Pass框架.
后面权重比较需要Depth, 这里先提前加了include DeclareDepthTexture.hlsl .
Shader &#34;HBAO&#34;
{
SubShader
{
...
Pass
{
Name &#34;HBAO&#34;
...
}
Pass
{
Name &#34;Blur&#34;
HLSLPROGRAM
#pragma vertex vert
#pragma fragment frag
#include &#34;Packages/com.unity.render-pipelines.core/ShaderLibrary/Common.hlsl&#34;
#include &#34;Packages/com.unity.render-pipelines.universal/ShaderLibrary/Core.hlsl&#34;
#include &#34;Packages/com.unity.render-pipelines.universal/ShaderLibrary/DeclareDepthTexture.hlsl&#34;
struct a2v
{
uint vertexID :SV_VertexID;
};
struct v2f
{
float4 pos:SV_Position;
float2 uv:TEXCOORD0;
};
v2f vert(a2v IN)
{
v2f o;
o.pos = GetFullScreenTriangleVertexPosition(IN.vertexID);
o.uv = GetFullScreenTriangleTexCoord(IN.vertexID);
return o;
}
half frag(v2f IN) : SV_Target
{
float2 uv = IN.uv;
return 0;
}
ENDHLSL
}
}
}
添加C#传入的属性 和 KERNEL_RADIUS(模糊半径).
Pass
{
Name &#34;Blur&#34;
HLSLPROGRAM
...
struct v2f
{
...
};
#define KERNEL_RADIUS 2
TEXTURE2D(_AOTex);
SAMPLER(sampler_AOTex);
float _BlurSharpness;
float2 _BlurDeltaUV;
v2f vert(a2v IN)
{
...
}
...
ENDHLSL
}
4.3 提前准备
在Blur之前还要写一些方法.
FetchAOAndDepth . 传入UV, 得到AO和线性的0~1深度.
...
float2 _BlurDeltaUV;
void FetchAOAndDepth(float2 uv, inout float ao, inout float depth)
{
ao = SAMPLE_TEXTURE2D_LOD(_AOTex, sampler_AOTex, uv, 0).r;
depth = SampleSceneDepth(uv);
depth = Linear01Depth(depth, _ZBufferParams);
}
v2f vert(a2v IN)
{
...
}
half frag(v2f IN) : SV_Target
{
...
}
再添加三个方法, CrossBilateralWeight, ProcessSample , ProcessRadius.
CrossBilateralWeight, 通过采样半径和深度差得到一个滤波权重.
ProcessSample, 累加 滤波权重 和 AO*滤波权重, 后面得到平均AO用.
ProcessRadius, 循环偏移UV进行采样和累加.
void FetchAOAndDepth(float2 uv, inout float ao, inout float depth)
{
...
}
float CrossBilateralWeight(float r, float d, float d0)
{
float blurSigma = KERNEL_RADIUS * 0.5;
float blurFalloff = 1.0 / (2.0 * blurSigma * blurSigma);
float dz = (d0 - d) * _ProjectionParams.z * _BlurSharpness;
return exp2(-r * r * blurFalloff - dz * dz);
}
void ProcessSample(float ao, float d, float r, float d0, inout float totalAO, inout float totalW)
{
float w = CrossBilateralWeight(r, d, d0);
totalW += w;
totalAO += w * ao;
}
void ProcessRadius(float2 uv0, float2 deltaUV, float d0, inout float totalAO, inout float totalW)
{
float ao;
float d;
float2 uv;
UNITY_UNROLL
for (int r = 1; r <= KERNEL_RADIUS; r++)
{
uv = uv0 + r * deltaUV;
FetchAOAndDepth(uv, ao, d);
ProcessSample(ao, d, r, d0, totalAO, totalW);
}
}
v2f vert(a2v IN)
{
...
}
4.4 Blur
直接在frag中调用刚才创建的方法. 即累加AO除权就好了.
half frag(v2f IN) : SV_Target
{
float2 uv = IN.uv;
float2 deltaUV = _BlurDeltaUV;
float totalAO;
float depth;
FetchAOAndDepth(uv, totalAO, depth);
float totalW = 1.0;
ProcessRadius(uv, -deltaUV, depth, totalAO, totalW);
ProcessRadius(uv, +deltaUV, depth, totalAO, totalW);
totalAO /= totalW;
return totalAO;
}
Blur这一步做完就是下图这样. 之后就是要Combine了.
<hr/>5. Combine
5.1 C#
最后就是把AO图贴回到主画布上.
我这里用的是Unity2022, API和2021有点不一样.
返回 HBAORenderPass.cs , 继续修改 Execute 方法. 其实就是把AOTex传给Shader, 再SetRT为ColorTarget, 最后执行全屏绘制.
public class HBAORenderPass : ScriptableRenderPass
{
...
public override void Execute(ScriptableRenderContext context, ref RenderingData renderingData)
{
var cmd = CommandBufferPool.Get();
using (new ProfilingScope(cmd, profilingSampler))
{
...
CoreUtils.DrawFullScreen(cmd, effectMat, null, 1);
cmd.SetGlobalTexture(aoTex_ID, hbaoRT_ID);
cmd.SetRenderTarget(renderingData.cameraData.renderer.cameraColorTargetHandle);
CoreUtils.DrawFullScreen(cmd, effectMat, null, 2);
cmd.ReleaseTemporaryRT(hbaoRT_ID);
...
}
context.ExecuteCommandBuffer(cmd);
CommandBufferPool.Release(cmd);
}
}
5.2 Shader
Shader这里准备用Color Blend(OM)模式来做, 我们把AO的值当作Alpha来输出. 最后就是 finalColor = DestRT.rgb(原图颜色) * srcAlpha(AO) + srcColor(白色) * 0 .
为什么原图的rgb要输出白色?
因为方便改Color Blend(OM)进行debug. 比如说 把 Blend SrcAlpha Zero 改成 Blend SrcAlpha Zero 就可以直接进行debug了.
修改 HBAO.Shader 文件, 加一个CombinePass.
Shader &#34;HBAO&#34;
{
SubShader
{
...
Pass
{
Name &#34;HBAO&#34;
...
}
Pass
{
Name &#34;Blur&#34;
...
}
Pass
{
Name &#34;Combine&#34;
Blend Zero SrcAlpha
// Blend SrcAlpha Zero
HLSLPROGRAM
#pragma vertex vert
#pragma fragment frag
#include &#34;Packages/com.unity.render-pipelines.core/ShaderLibrary/Common.hlsl&#34;
#include &#34;Packages/com.unity.render-pipelines.universal/ShaderLibrary/Core.hlsl&#34;
struct a2v
{
uint vertexID :SV_VertexID;
};
struct v2f
{
float4 pos:SV_Position;
float2 uv:TEXCOORD0;
};
TEXTURE2D(_AOTex);
SAMPLER(sampler_AOTex);
v2f vert(a2v IN)
{
v2f o;
o.pos = GetFullScreenTriangleVertexPosition(IN.vertexID);
o.uv = GetFullScreenTriangleTexCoord(IN.vertexID);
return o;
}
half4 frag(v2f IN) : SV_Target
{
half ao = SAMPLE_TEXTURE2D_LOD(_AOTex, sampler_AOTex, IN.uv, 0).x;
return half4(1, 1, 1, ao);
}
ENDHLSL
}
}
}
这样做完了一个简单的HBAO. 下面是Combine之后的对比图, 效果能凑活着用.
<hr/>在计算AO和Blur阶段, 感觉用Compute Shader应该会快一点. 用groupshared 去缓存depth, 从而减少采样.
项目中感觉可以用半分辨率+更少的方向划分和步进, 效果其实差不了多少.
下面还有一些别人的文章, 里面有代码和其它的AO可以做参考.
图形学基础|环境光遮蔽（Ambient Occlusion）
【论文复现】Image-Space Horizon-Based Ambient Occlusion
环境遮罩
<hr/>写完印度の剑3正好上线, 芜湖起飞! |
本帖子中包含更多资源
您需要 登录 才可以下载或查看，没有账号？立即注册
×
|
窥视卡
雷达卡
发表于 2022-8-4 19:57
提升卡
置顶卡
沉默卡
喧嚣卡
变色卡
千斤顶
显身卡