上海
【USparkle專欄】如果你深懷絕技，愛“搞點研究”，樂于分享也博采眾長，我們期待你的加入，讓智慧的火花碰撞交織，讓知識的傳遞生生不息！
這是侑虎科技第1939篇文章，感謝作者傻頭傻腦亞古獸供稿。歡迎轉發分享，未經作者授權請勿轉載。如果您有任何獨到的見解或者發現也歡迎聯系我們，一起探討。（QQ群：793972859）
作者主頁：
https://www.zhihu.com/people/tian-cai-ya-gu-shou
一、前序
1. 介紹
Nanite是UE5中虛擬幾何體（Virtualized Geometry System）的系統，主要用途是高效率渲染的高面數模型。Nanite會為模型自動生成LOD結構，與傳統LOD不同，Nanite的LOD不再是每個模型的，而是精細到模型中的局部區域，藝術家不需再為制作或處理LOD煩惱。并且還能享有GPU Driven的高效剔除，單個繪制調用的好處。
2. 技術要點
Nanite技術結合了多種技術做到了高效渲染：
1. Cluster Rendering：由Cluster組織三角形，可以享有更高效的剔除。
2. Auto LOD：通過Graph Partitioning技術劃分和簡化模型構建LOD，并且把數據組織成BVH結構在Runtime時候可以高效地并行選擇LOD，通過這種方式構建的LOD過渡非常絲滑。
3. GPU Driven Pipeline：由GPU驅動的繪制，減少了CPU的性能開銷。
4. Occlusion Culling：更細顆粒的遮擋剔除，用于剔除不可見的三角形。
5. Hardware/Software Rasterization：由于小三角形對于硬件光柵化非常不友好，所以針對這些三角形用Compute Shader執行軟光柵提高效率。
6. Visibility Buffer：利用Visibility Buffer減少Overdraw，進一步提高GPU效率。
7. Streaming：加載只看到的相關數據，減少幾何體對內存的壓力。
3. 本文效果
由于Nanite系統非常龐大和有非常多的工程細節要處理，所以本文會簡化和略過一些東西，僅實現核心部分，而且會與有UE5的版本有點出入。
下圖是本文實現的效果，每個色塊是一個三角形，可以看出LOD切換和相機剔除都非常絲滑。
色塊表示三角面
色塊表示Cluster
二、實現
1. Clusterize
第一步，在離線階段處理，將復雜的超高精度網格模型高效且合理地分割成更小、更易于管理的簇（Cluster），每個Cluster最多128個三角形。這種劃分不是簡單的切割，而是旨在最小化簇與簇之間連接的邊數（即切割大小），同時保持每個簇的大小大致均衡。
UE使用的Partition是Metis庫：
https://github.com/KarypisLab/METIS
實現代碼可以參考UE5的源碼部分：
UnrealEngine-release\Engine\Source\Developer\NaniteBuilder\Private\NaniteBuilder.cpp
本文使用meshoptimizer實現Mesh的切分Cluster和Partition功能，這個庫功能還有優化Over Draw，Shadow Depth Index等功能：
https://github.com/zeux/meshoptimizer
我們新建一個C++導出DLL的工程，封裝幾個主要函數讓Unity可以使用。其實代碼量不多，翻譯成C# 直接用也可以。
分別是：
meshopt_buildMeshlets（構建Cluster）
meshopt_partitionClusters（Cluster劃分Partition）
meshopt_buildMeshletsBound（計算Cluster數量）
meshopt_computeSphereBounds（合并BoundsSphere）
在C# 中引用這些函數：
unsafe static List clusterize(Vector3[] vertices, int[] indices)
{
constint max_vertices = 192; // TODO: depends on kClusterSize, also may want to dial down for mesh shaders
constint max_triangles = kClusterSize; //128
constint min_triangles = (kClusterSize / 3) & ~3;
constfloat split_factor = 2.0f;
constfloat fill_weight = 0.75f;
int max_meshlets = BuildMeshletsBound(indices.Length, max_vertices, max_triangles);//meshopt_buildMeshletsBound
var meshlets = new Meshlet[max_meshlets * 2];
var meshlet_vertices = newint[max_meshlets * max_vertices];
var meshlet_triangles = newbyte[max_meshlets * max_triangles * 3];
var meshlet_count = BuildMeshletFlex(meshlets, meshlet_vertices, meshlet_triangles, indices, indices.Length, vertices, vertices.Length, sizeof(float) * 3, max_vertices, min_triangles, max_triangles, 0.0f,
split_factor);//meshopt_buildMeshlets
List clusters = new List (meshlet_count);
for (int i = 0; i < meshlet_count; i++)
{
ref Meshlet meshlet = ref meshlets[i];
fixed (int* ptr = &meshlet_vertices[meshlet.vertex_offset])
{
fixed (byte* ptr2 = &meshlet_triangles[meshlet.triangle_offset])
{
OptimizeMeshlet(ptr, ptr2, (int)meshlet.triangle_count, (int)meshlet.vertex_count);
}
}
Cluster cluster = new Cluster();
cluster.indices = newint[meshlet.triangle_count * 3];
for (int j = 0; j < meshlet.triangle_count * 3; ++j)
cluster.indices[j] =
meshlet_vertices[meshlet.vertex_offset + meshlet_triangles[meshlet.triangle_offset + j]];
cluster.parent.error = float.MaxValue;
clusters.Add(cluster);
}return clusters;
}
然后可以直接通過meshopt_buildMeshlets函數，獲得每個cluster的indexs。
2. Build DAG
有了這些Cluster，就可以構建“LOD”了，只需要循環這個操作：打組->合并->減面->clusterize。如下圖：
這個過程感覺就像Mipmap一樣，一層一層往上合并和簡化，并記錄一個Err誤差值和Bounds用于運行時LOD選擇用。而這些合并的的節點就叫做Cluster Group。最后得出一個DAG（有向無環圖，Directed Acyclic Graph）的結構。
public struct ClusterGroup
{
public List Children;
public Vector3 Bounds;
publicfloat radius;
public Vector3 LODBounds;
publicfloat MinLODError;
publicfloat MaxParentLODError;
publicint MipLevel;
}
publicclassNaniteSubMesh
{
public List Group> clusterGroupList;
public List clusterList;
publicint maxMipLevel;
}
static NaniteSubMesh Nanite(Vector3[] vertices,Vector3[] normals, int[] indices)
{
NaniteSubMesh res = new NaniteSubMesh();
List Group> clusterGroupList = new List Group>();
var clusters = clusterize(vertices, indices);
res.clusterList = clusters;
res.clusterGroupList = clusterGroupList;
res.maxMipLevel = 0;
for (int i = 0; i < clusters.Count; ++i)
{
var c = clusters[i];
c.self = Bounds(vertices, clusters[i].indices, 0f);
c.mip = 0;
clusters[i] = c;
}
List pending = new List(clusters.Count);
int[] remap = newint[vertices.Length];
for (int i = 0; i < remap.Length; ++i)
remap[i] = i;
for (int i = 0; i < clusters.Count; ++i)
pending.Add(i);
int curMip = 1;
byte[] locks = newbyte[vertices.Length];
while (pending.Count > 1)
{
List int>> groups = partition(clusters, pending, remap, vertices);
if (kUseLocks)
lockBoundary(locks, groups, clusters, remap);
pending.Clear();
List retry = new List();
int triangles = 0;
int stuck_triangles = 0;
for (int i = 0; i < groups.Count; ++i)
{
var curGroupClusters = groups[i];
if (curGroupClusters.Count == 0)
{
continue; // metis shortcut
}
List merged = new List(vertices.Length);
for (int j = 0; j < curGroupClusters.Count; ++j)
{
merged.AddRange(clusters[curGroupClusters[j]].indices);
}
LODBounds groupb = boundsMerge(clusters, curGroupClusters);
ClusterGroup clusterGroup = new ClusterGroup();
clusterGroup.Bounds = groupb.center;
clusterGroup.MaxParentLODError = groupb.error;
clusterGroup.radius = groupb.radius;
clusterGroup.Children = new List(merged.Count);
clusterGroup.MipLevel = curMip - 1;
for (int j = 0; j < curGroupClusters.Count; ++j)
{
clusterGroup.Children.Add(curGroupClusters[j]);
}
clusterGroupList.Add(clusterGroup);
// aim to reduce group size in half
int target_size = (merged.Count / 3) / 2 * 3;
float error = 0f;
var simplified = simplify(vertices, normals, merged.ToArray(), kUseLocks ? locks : null, target_size,
ref error);
if (simplified.Count > merged.Count * kSimplifyThreshold)
{
stuck_triangles += merged.Count / 3;
for (int j = 0; j < curGroupClusters.Count; ++j)
{
retry.Add(curGroupClusters[j]);
}
continue; // simplification is stuck; abandon the merge
}
// enforce bounds and error monotonicity
// note: it is incorrect to use the precise bounds of the merged or simplified mesh, because this may violate monotonicity
var split = clusterize(vertices, simplified.ToArray());
groupb.error += error; // this may overestimate the error, but we are starting from the simplified mesh so this is a little more correct
// update parent bounds and error for all clusters in the group
// note that all clusters in the group need to switch simultaneously so they have the same bounds
for (int j = 0; j < curGroupClusters.Count; ++j)
{
int clusterIndex = curGroupClusters[j];
var t = clusters[clusterIndex];
t.parent = groupb;
clusters[clusterIndex] = t;
}
for (int j = 0; j < split.Count; ++j)
{
var sj = split[j];
sj.self = groupb;
sj.mip = curMip;
split[j] = sj;
clusters.Add(sj); // std::move
pending.Add(clusters.Count - 1);
triangles += sj.indices.Length / 3;
}
}
curMip++;
}
if (pending.Count == 1)
{
var c = clusters[pending[0]];
ClusterGroup clusterGroup = new ClusterGroup();
clusterGroup.Bounds = c.self.center;
clusterGroup.MaxParentLODError = c.self.error;
clusterGroup.radius = c.self.radius;
clusterGroup.Children = new List(1);
clusterGroup.MipLevel = curMip - 1;
clusterGroup.Children.Add(pending[0]);
clusterGroupList.Add(clusterGroup);
}
res.maxMipLevel = curMip - 1;
return res;
}
static void lockBoundary(byte[] locks, List int>> groups, List clusters, int[] remap)
{
// for each remapped vertex, keep track of index of the group it's in (or -2 if it's in multiple groups)
int[] groupmap = newint[locks.Length];
for (int i = 0; i < groupmap.Length; ++i)
groupmap[i] = -1;
for (int i = 0; i < groups.Count; ++i)
{
var c = groups[i];
for (int j = 0; j < c.Count; ++j)
{
var indices = clusters[c[j]].indices;
for (int k = 0; k < indices.Length; ++k)
{
var v = indices[k];
var r = remap[v];
if (groupmap[r] == -1 || groupmap[r] == i)
groupmap[r] = i;
else
groupmap[r] = -2;
}
}
}// note: we need to consistently lock all vertices with the same position to avoid holes
for (int i = 0; i < locks.Length; ++i)
{
var r = remap[i];
locks[i] = (byte)((groupmap[r] == -2) ? 1 : 0);
}
}
這樣我們得到各級Mip的一系列Clusters。
3. 加速結構
即使把三角形劃分成Clusters數量也太多，使用Compute Shader來做并行結算效率也不高，于是Nanite就使用了BVH來作為ClusterGroup的加速結構，然后配合Persistent Threads做查找過濾。
Persistent Threads遍歷BVH部分，有興趣可以參考UE5源碼：Shaders\Private\Nanite\NaniteClusterCulling.usf
UE5中也有不使用Persistent Threads的流程，應該說一般默認就是不使用的。
UE5源碼部分
個人認為Persistent Threads方案在GPU遍歷這種BVH結構有點暴力和重度，所以簡化了一下，把多個Cluster合并成一個剔除單元（Part），先并行對Part做剔除，再對Part里的Cluster去做并行剔除，兩層結構來加速作為Persistent Threads的一個簡單替代方案。
然后把多個Part組織成Page用于分塊加載。材質處理細節也不同，UE5的材質是每個Cluster會記錄MaterialRange，簡單起見這里實現是每個SubMesh會去構建獨立的Clusters。
代碼如下：
[Serializable]
publicstruct NaniteCluster
{
publicint indiceIndex;
publicint indiceCount;
publicfloat selfErrer;
publicfloat parentErrer;
public Vector4 selfSphere;
public Vector4 parentSphere;
publicint subMeshID;
publicint vertexOffset;
};
[Serializable]
publicstruct NaniteClusterGroup
{
publicint ClusterStart;
publicint ClusterCount;
public Vector3 Bounds;
publicfloat radius;
public Vector3 LODBounds;
publicfloat MinLODError;
publicfloat MaxParentLODError;
publicint MipLevel;
}[Serializable]
publicstruct NaniteMeshPart
{
publicint ClusterStart;
publicint ClusterCount;
public Vector4 selfSphere;
publicfloat MaxParentLODError;
}
public classNaniteSubMesh
{
public List Group> clusterGroupList;
public List clusterList;
publicint maxMipLevel;
}
publicclassBuildPart
{
public List clusterList;
publicint mip;
publicint subMesh;
}
public static void BuildNaniteMesh(Mesh mesh)
{
var vertices = mesh.vertices;
var normals = mesh.normals;
var uvs = mesh.uv;
int subMeshCount = mesh.subMeshCount;
int totalClusterCount = 0;
int totalIndexCount = 0;
List subMeshList = new List ();
for (int i = 0; i < subMeshCount; i++)
{
var triangles = mesh.GetTriangles(i);
var subMesh = Nanite(vertices,normals,triangles);
subMeshList.Add(subMesh);
totalClusterCount += subMesh.clusterList.Count;
}
List buildPartsList = new List (totalClusterCount);
int MAX_PART_PERPAGE = 128;
int MAX_CLUSTER_PERPART = 8;
for (int subMeshIndex = 0; subMeshIndex < subMeshList.Count; subMeshIndex++)
{
var subMesh = subMeshList[subMeshIndex];
List clusters = subMesh.clusterList;
var groupsList = subMesh.clusterGroupList;
BuildPart buildPart = null;
for (int i = 0; i < groupsList.Count; i++)
{
var gIndex = i; // sortGroups[i].OldIndex;
var g = groupsList[gIndex];
var childs = g.Children;
for (int c = 0; c < childs.Count; c++)
{
int cIndex = childs[c];
int cMip = clusters[cIndex].mip;
totalIndexCount += clusters[cIndex].indices.Length;
//new Part
if (buildPart == null || buildPart.clusterList.Count >= MAX_CLUSTER_PERPART ||
buildPart.mip != cMip)
{
buildPart = new BuildPart();
buildPart.clusterList = new List(MAX_CLUSTER_PERPART);
buildPart.mip = cMip;
buildPart.subMesh = subMeshIndex;
buildPartsList.Add(buildPart);
}
buildPart.clusterList.Add(cIndex);
}
}
}
int buildPartCount = buildPartsList.Count;
NaniteMeshPage[] pageArray = new NaniteMeshPage[(buildPartCount+(MAX_PART_PERPAGE-1))/MAX_PART_PERPAGE];//ceil
List tempIndiceList = new List(totalIndexCount);
List mipLists = new List(totalClusterCount);
int partIndex = 0;
for (int i = 0; i < pageArray.Length; i++)
{
//create new page
var p = ScriptableObject.CreateInstance ();
pageArray[i] = p;
tempIndiceList.Clear();
int partCount = (i == (pageArray.Length -1)) ? (buildPartCount % MAX_PART_PERPAGE) : MAX_PART_PERPAGE;
p.parts = new NaniteScene.NaniteMeshPart[partCount];
List pageClusters = new List (partCount * MAX_CLUSTER_PERPART);
for (int j = 0; j < partCount; j++)
{
var buildPart = buildPartsList[partIndex];
var buildPartCluster = buildPart.clusterList;
//create part
var part = new NaniteScene.NaniteMeshPart();
part.ClusterStart = pageClusters.Count; //local index
part.ClusterCount = buildPartCluster.Count;
int subMeshID = buildPart.subMesh;
float maxParentErr = 0f;
var clusters = subMeshList[subMeshID].clusterList;
for (int c = 0; c < buildPartCluster.Count; c++)
{
var cluster = clusters[buildPartCluster[c]];
mipLists.Add(cluster.mip);
//create Cluster
NaniteScene.NaniteCluster naniteCluster = new NaniteScene.NaniteCluster();
naniteCluster.indiceIndex = tempIndiceList.Count;
naniteCluster.indiceCount = cluster.indices.Length;
naniteCluster.parentErrer = cluster.parent.error;
naniteCluster.parentSphere = new Vector4(cluster.parent.center.x,cluster.parent.center.y,cluster.parent.center.z, cluster.parent.radius);
naniteCluster.selfErrer = cluster.self.error;
naniteCluster.selfSphere = new Vector4(cluster.self.center.x,cluster.self.center.y,cluster.self.center.z, cluster.self.radius);
naniteCluster.subMeshID = subMeshID;
tempIndiceList.AddRange(cluster.indices);
maxParentErr = Mathf.Max(naniteCluster.parentErrer, maxParentErr);
pageClusters.Add(naniteCluster);
}
LODBounds partBounds = boundsMerge(clusters, buildPartCluster,true);
part.selfSphere = new Vector4(partBounds.center.x,partBounds.center.y,partBounds.center.z,partBounds.radius);
part.MaxParentLODError = maxParentErr;
p.parts[j] = part;
partIndex++;
}
p.clusterArray = pageClusters.ToArray();
p.indiceArray = tempIndiceList.ToArray();
p.clusterMip = mipLists.ToArray();
}
string fileName = AssetDatabase.GetAssetPath(mesh);
string extension = Path.GetExtension(fileName);
fileName = fileName.Replace(extension, "");
//Build page
int totalVerts = 0;
for (int i = 0; i < pageArray.Length; i++)
{
var page = pageArray[i];
var clusterArray = page.clusterArray;
var indiceArray = page.indiceArray;
Dictionary indicesMap = new Dictionary();
List tempVerts = new List (vertices.Length);
List tempNormals = new List (vertices.Length);
List tempUVs = new List (vertices.Length);
List newIndices = new List(totalIndexCount);
for (int c = 0; c < clusterArray.Length; c++)
{
refvar cluster = ref clusterArray[c];
var indexStart = cluster.indiceIndex;
var indexEnd = indexStart+cluster.indiceCount;
for (int index = indexStart; index < indexEnd; index++)
{
int vertIndex = indiceArray[index];
int newIndex;
if (!indicesMap.TryGetValue(vertIndex,out newIndex))
{
newIndex = newIndices.Count;
indicesMap.Add(vertIndex, newIndex);
tempVerts.Add(vertices[vertIndex]);
tempNormals.Add(normals[vertIndex]);
if (uvs.Length == 0)
{
tempUVs.Add(Vector2.zero);
}
else
{
tempUVs.Add(uvs[vertIndex]);
}
newIndices.Add(newIndex);
}
indiceArray[index] = newIndex;
}
}
page.vertexStride = 5;//pos3 + uv2
page.vertexData = newfloat[tempVerts.Count * page.vertexStride];
page.vertexCount = tempVerts.Count;
for (int v = 0; v < tempVerts.Count; v++)
{
int vertexIndex = v * page.vertexStride;
page.vertexData[vertexIndex + 0] = tempVerts[v].x;
page.vertexData[vertexIndex + 1] = tempVerts[v].y;
page.vertexData[vertexIndex + 2] = tempVerts[v].z;
page.vertexData[vertexIndex + 3] = tempUVs[v].x;
page.vertexData[vertexIndex + 4] = tempUVs[v].y;
}
totalVerts +=tempVerts.Count;
string newPath = fileName + "_p"+i +".asset";
AssetDatabase.CreateAsset(page, newPath);
}
AssetDatabase.Refresh();
Debug.Log("mesh Vertx:"+vertices.Length +" mesh Nanite:"+ totalVerts + " cluster:"+totalClusterCount + "part:"+ buildPartCount +" page:"+pageArray.Length);
NaniteMesh naniteMesh = ScriptableObject.CreateInstance ();
{
naniteMesh.subMeshCount = subMeshCount;
naniteMesh.pageArray = new NaniteMeshPage[pageArray.Length];
for (int i = 0; i < pageArray.Length; i++)
{
string newPath = fileName + "_p" + i + ".asset";
naniteMesh.pageArray[i] = AssetDatabase.LoadAssetAtPath (newPath);
}
}var meshBound = mesh.bounds;
naniteMesh.boundingSphere = meshBound.center;
naniteMesh.boundingSphere.w = meshBound.extents.magnitude;
string meshExt = "_mesh.asset";
AssetDatabase.CreateAsset(naniteMesh, fileName + meshExt);
AssetDatabase.Refresh();
}
到這里離線部分基本結束，可以得到一個Nanite的資源。當然UE5原文還做了很多操作，如BVH、Encode、編碼、壓縮、Page的劃分、頂點屬性優化等，個人認為這些都屬于工程細節。
4. 運行時資源
來到Runtime部分，我們需要把這個Nanite Mesh加載上來，方便起見，這里直接引用一下資源在腳本上，偷懶省略加載部分。
把資源、Object、材質信息整合起來，傳到GPU的Buffer中。這里做法很不正式還是偷懶來處理。當然也可以用Compute Shader來更新Page數據到GPUBuffer中。
public static List renderers = new List ();
privatestatic SceneObject[] gpuObjects = new SceneObject[2048];
//cluster -> part -> page
publicstruct SceneObject
{
publicint naniteMeshID;
public Matrix4x4 localToWorldMatrix;
publicint materialIDOffset;
}
publicstruct NaniteRes
{
public Vector4 boundingSphere;
publicint partIndex;
publicint partCount;
}
unsafe static void UpdateRenderList()
{
if(renderers.Count == 0)
return;
//object update
if (renderers.Count > gpuObjects.Length)
{
gpuObjects = new SceneObject[Mathf.NextPowerOfTwo(renderers.Count)];
}
objectCount = 0;
maxPartCount = 0;
naniteMeshes.Clear();
materialList.Clear();
List materialIndices = new List();
for (int i = 0; i < renderers.Count; i++)
{
var renderer = renderers[i];
var nMesh = renderer.naniteMesh;
foreach (var p in nMesh.pageArray)
{
maxPartCount += p.parts.Length;
maxClusterCount += p.clusterArray.Length;
}
SceneObject obj = new SceneObject();
obj.localToWorldMatrix = renderer.transform.localToWorldMatrix;
//mesh index
int index = naniteMeshes.IndexOf(nMesh);
if (index < 0)
{
index = naniteMeshes.Count;
naniteMeshes.Add(nMesh);
}
obj.naniteMeshID = index;
//mat indexs
obj.materialIDOffset = materialIndices.Count;
for (int m = 0; m < renderer.materials.Length; m++)
{
var mat = renderer.materials[m];
int matIndex = materialList.IndexOf(mat);
if (matIndex < 0)
{
matIndex = materialList.Count;
materialList.Add(mat);
}
materialIndices.Add(matIndex);
}
gpuObjects[i] = obj;
renderer.transformChanged = false;
objectCount++;
}
if(candidateClusterBuffer!=null)
candidateClusterBuffer.Dispose();
candidateClusterBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Structured, maxClusterCount *2, sizeof(int));
if(visibleClusterBuffer != null)
visibleClusterBuffer.Dispose();
visibleClusterBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Structured,maxClusterCount *2, sizeof(int));
if (objectsBuffer != null)
objectsBuffer.Dispose();
objectsBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Structured, objectCount, sizeof(SceneObject));
objectsBuffer.SetData(gpuObjects,0,0,objectCount);
if(visObjectsBuffer !=null)
visObjectsBuffer.Dispose();
visObjectsBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Structured,objectCount, sizeof(int));
int vertCount = 0;
List tempClusters = new List ( 2048);
List tempParts = new List ( 2048);
List naniteRes = new List ( 2048);
List tempIndices = new List(2048 * 100);
List vertexDataList = new List();
//load page
for (int nID = 0; nID < naniteMeshes.Count; nID++)
{
NaniteRes res = new NaniteRes();
var nMesh = naniteMeshes[nID];
//填充到GPU
var pages = nMesh.pageArray;
res.partIndex = tempParts.Count;
res.partCount = 0;
res.boundingSphere = nMesh.boundingSphere;
for (int p = 0; p < pages.Length; p++)
{
var page = pages[p];
var parts = page.parts;
int vertOffset = vertCount;
int indicesOffset = tempIndices.Count;
int clusterOffset = tempClusters.Count;
//add all cluster
var clusters = page.clusterArray;
for (int c = 0; c < clusters.Length; c++)
{
var cluster = clusters[c];
cluster.indiceIndex += indicesOffset;
cluster.vertexOffset = vertOffset;
tempClusters.Add(cluster);
}
//add all part
for (int partIndex = 0; partIndex < parts.Length; partIndex++)
{
var part = parts[partIndex];
part.ClusterStart += clusterOffset;
tempParts.Add(part);
res.partCount++;
}
//add page data
tempIndices.AddRange( page.indiceArray);
vertexDataList.AddRange(page.vertexData);
vertCount += page.vertexCount;
}
naniteRes.Add(res);
}
//TODO GPU Update Buffer
if (naniteResBuffer != null)
naniteResBuffer.Dispose();
naniteResBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Structured, naniteRes.Count, sizeof(NaniteRes));
naniteResBuffer.SetData(naniteRes);
if (partsBuffer != null)
partsBuffer.Dispose();
partsBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Structured,tempParts.Count, sizeof(NaniteMeshPart));
partsBuffer.SetData(tempParts);
if (clusterBuffer != null)
clusterBuffer.Dispose();
clusterBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Structured, tempClusters.Count, sizeof(NaniteCluster));
clusterBuffer.SetData(tempClusters);
if (indiceseBuffer != null)
indiceseBuffer.Dispose();
indiceseBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Raw, tempIndices.Count, sizeof(int));
indiceseBuffer.SetData(tempIndices);
if(materialIndexBuffer!=null)
materialIndexBuffer.Dispose();
materialIndexBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Structured,materialIndices.Count, sizeof(int));
materialIndexBuffer.SetData(materialIndices);
if(vertexDataBuffer!=null)
vertexDataBuffer.Dispose();
vertexDataBuffer = new GraphicsBuffer(GraphicsBuffer.Target.Raw, vertexDataList.Count,sizeof(float));
vertexDataBuffer.SetData(vertexDataList);
}
//input object ID =>
public unsafe static void UpdateNaniteScene()
{
if (renderListDirty)
{
UpdateRenderList();
// UpdateRenderListGPU();
renderListDirty = false;
}
for (int i = 0; i < renderers.Count; i++)
{
var renderer = renderers[i];
if (renderer.transformChanged)
{
gpuObjects[i].localToWorldMatrix = renderer.transform.localToWorldMatrix;
renderer.transformChanged = false;
transformDirty = true;
}
}if (objectsBuffer != null && transformDirty)
objectsBuffer.SetData(gpuObjects, 0, 0, objectCount);
}
5. 剔除
這時離線時候已經把Clusters扁平化到數組中了，這些Clusters是可以并行進行剔除的，巧妙之處是他記錄了父級的誤差和自己的誤差，當我們傳入誤差系數時候就可以獨立地判斷自己是否被剔除，而和上下級無關。
先從CPU發起剔除Compute Shader的Dispatch。這里因為組織數據時候就知道了所有Object最大的Parts/Cluster數量，所以直接用這個數去Dispatch了。
Objects剔除：
根據Object找到NaniteMesh的Parts進行Culling：
ClustersCulling：
6. 軟光柵
略。
7. VisibilityBuffer
VBuffer主要用來減少Overdraw，著色器直接輸出InstanceID、ClusterID、材質ID。然后用這個VBuffer來計算頂點數據來著色。
這個得益于GPUDriven的好處，一個DrawProceduralIndirect就可以繪制所有物體了：
一次DrawProceduralIndirect繪制多個物體
VBuffer存哪些屬性，多少位，都是工程細節這里就不考究了。
8. 著色
有了VBuffer就需要逐材質進行繪制，原文是材質ID分Tile組合IndirectDraw畫Quad的思想。
需要注意一下這里VBuffer通過三角重心插值求出的UV是不能直接采樣貼圖的，因為DDXY不對，所以需求重新計算，計算的代碼放下面。并且利用SampleGrad（samplerName, coord2, dpdx, dpdy）來采樣。
uint MurmurMix(uint Hash)
{
Hash ^= Hash >> 16;
Hash *= 0x85ebca6b;
Hash ^= Hash >> 13;
Hash *= 0xc2b2ae35;
Hash ^= Hash >> 16;
return Hash;
}
float3 IntToColor(uint Index)
{
uint Hash = MurmurMix(Index);
float3 Color = float3
(
(Hash >> 0) & 255,
(Hash >> 8) & 255,
(Hash >> 16) & 255
);
return Color * (1.0f / 255.0f);
}
struct FBarycentrics
{
float3 Value;
float3 Value_dx;
float3 Value_dy;
};
float2 Lerp(float2 Value0, float2 Value1, float2 Value2, FBarycentrics Barycentrics, out float2 dxy)
{
float2 Value = Value0 * Barycentrics.Value.x + Value1 * Barycentrics.Value.y + Value2 * Barycentrics.Value.z;
dxy.x = Value0 * Barycentrics.Value_dx.x + Value1 * Barycentrics.Value_dx.y + Value2 * Barycentrics.Value_dx.z;
dxy.y = Value0 * Barycentrics.Value_dy.x + Value1 * Barycentrics.Value_dy.y + Value2 * Barycentrics.Value_dy.z;
return Value;
}
/** Calculates perspective correct barycentric coordinates and partial derivatives using screen derivatives. */
FBarycentrics CalculateTriangleBarycentrics(float2 PixelClip, float4 PointClip0, float4 PointClip1,
float4 PointClip2, float2 ViewInvSize)
{
FBarycentrics Barycentrics;
PixelClip.y = 1 - PixelClip.y;
PixelClip.xy = PixelClip.xy * 2 - 1;
const float3 RcpW = rcp(float3(PointClip0.w, PointClip1.w, PointClip2.w));
const float3 Pos0 = PointClip0.xyz * RcpW.x;
const float3 Pos1 = PointClip1.xyz * RcpW.y;
const float3 Pos2 = PointClip2.xyz * RcpW.z;
const float3 Pos120X = float3(Pos1.x, Pos2.x, Pos0.x);
co...
特別聲明：以上內容(如有圖片或視頻亦包括在內)為自媒體平臺“網易號”用戶上傳并發布，本平臺僅提供信息存儲服務。
Notice: The content above (including the pictures and videos if any) is uploaded and posted by a user of NetEase Hao, which is a social media platform and only provides information storage services.
