2023.02.12

• 更加快速的CPU视锥剔除
• Nanite Occlusion Culling
• C++ 单例

更加快速的CPU视锥剔除

GPU除外，CPU的frustum culling还有什么优化的空间
https://www.zhihu.com/question/313499217/answer/2862700086

• 缓存友好
  所有bounding相关信息存到一个数组里

    /**
   * Bounding information used to cull primitives in the scene.
   */
  struct FPrimitiveBounds
  {
  	FBoxSphereBounds BoxSphereBounds;
  	/** Square of the minimum draw distance for the primitive. */
  	float MinDrawDistanceSq;
  	/** Maximum draw distance for the primitive. */
  	float MaxDrawDistance;
  	/** Maximum cull distance for the primitive. This is only different from the MaxDrawDistance for HLOD.*/
  	float MaxCullDistance;
  };
  /** Packed array of primitive bounds. */
  TArray<FPrimitiveBounds> PrimitiveBounds;

• 多线程
• SIMD

  // Process four planes at a time until we have < 4 left
  for (int Count = 0; Count < Planes.Num(); Count += 4)
  {
  	// Load 4 planes that are already all Xs, Ys, ...
  	Register128 PlanesX = SIMD_Load(PlanePtr);
  	FrustumPlanesPlanesPtr++;
  	Register128 PlanesY = SIMD_Load(FrustumPlanesPlanesPtr);
  	FrustumPlanesPlanesPtr++;
  	Register128 PlanesZ = SIMD_Load(FrustumPlanesPlanesPtr);
  	FrustumPlanesPlanesPtr++;
  	Register128 PlanesW = SIMD_Load(FrustumPlanesPlanesPtr);
  	FrustumPlanesPlanesPtr++;
  
  	// Calculate the distance (x * x) + (y * y) + (z * z) - w
  	Register128 DistX = SIMD_Mul(OrigX,PlanesX);
  	Register128 DistY = SIMD_Mad(OrigY,PlanesY,DistX);
  	Register128 DistZ = SIMD_Mad(OrigZ,PlanesZ,DistY);
  	Register128 Distance = SIMD_Sub(DistZ,PlanesW);
  	
  	// Check for completely outside
  	if (SIMD_GreaterThan(Distance,VRadius))
  	{å
  		return false;
  	}
  }

  GDC16: Improving geometry culling for Deus Ex: Mankind Divided by Nicolas Trudel

Nanite Occlusion Culling

https://zhuanlan.zhihu.com/p/583245401

• Software Occulsion Culling
  https://www.intel.com/content/www/us/en/developer/articles/technical/software-occlusion-culling.html
  我们现在用的是occluder调用 Intel的Software Occulsion Culling
  • 用CPU对occluder做软光栅化生成depth buffer
  • 对所有物体做depth test culling
• gpu driven管线则可以把frustum culling, occlusion culling, cluster draw 全都放到GPU上
• 大革命
  • 提出假设，两帧画面的zbuffer具有足够的连续性，因此我们可以把上一帧的zbuffer投影到当前帧进行遮挡剔除
    • 遮挡问题
    • 动态物体 延迟
  • two pass
    
• Nanite
  • two pass 1
    
  • two pass 2
    • 这一帧可见的cluster大概率在上一帧也可见
    • 把当前帧所有cluster都转换到上一帧的位置上去，然后利用上一帧的hi-z进行遮挡剔除
    • 再对被剔除的cluster进行第二次保守剔除

C++ 单例