GPU Instancing及其应用

创建时间：2021/5/17 21:44

• GPU Instancing及其应用
  • 相关知识点
    • ComputeBuffer
      • Append-consume ComputeBuffer
      • ComputeBuffer.CopyCount
      • Other ComputeBuffer
  • Graphics:DrawMeshInstanced
    • DrawMeshInstanced
    • DrawMeshInstancedIndirect
    • DrawMeshInstancedProcedural
    • Benchmark
      • 不同方式绘制的性能
      • 关于剔除

相关知识点

ComputeBuffer

心得：

• ComputeBuffer种类多，操作相关的API类型多，其根本原因在于，GPU的智能化程度低，CPU去访问现存操作数据不方便，效率低，因此，指令种类复杂，可定制程度低。

• ComputeBuffer可以由ComputeShader使用，也可以有Shader使用，作为 ComputeShader与常规渲染Shader的桥梁，使得两者之间传递信息的效率较高，Shader功能的可定制性、可扩展性增强，ComputeShader与Shader可以灵活地结合使用。

• ComputeShader的优势在于

  • GPU计算，擅长处理大量并行化的任务，效率高

  • 计算得到的结果不需要用CPU获取，可以直接用于渲染，数据传递效率高。

Append-consume ComputeBuffer

• AppendStructuredBuffer - Output buffer that appears as a stream the shader may append to.

• ConsumeStructuredBuffer - An input buffer that appears as a stream the shader may pull values from.

ComputeBuffer.CopyCount

Copy counter value of append/consume buffer into another buffer.

Append/consume and counter buffers keep track of the number of elements in them with a special counter variable. CopyCount takes a buffer as src, and copies its counter value into dst buffer at given byte offset.

This is most commonly used in conjunction with Graphics.DrawProceduralIndirect, to render arbitrary number of primitives without reading their count back to the CPU.

On DX11 there is a restriction on the dst buffer - it must have been created with ComputeBufferType.Raw or ComputeBufferType.IndirectArguments type. On other platforms dst can be any type.

Other ComputeBuffer

Graphics:DrawMeshInstanced

DrawMeshInstanced

draws meshes for one frame without the overhead of creating unnecessary game objects.

用处： Use this function in situations where you want to draw the same mesh for a particular amount of times using an instanced shader.

包围盒计算： It creates an axis-aligned bounding box that contains all the Meshes , calculates the center point, then uses this information to cull and sort the Mesh instances.

关于剔除： Unity culls and sorts instanced Meshes as a group. 粗粒度剔除：对于 combined instances ，要么 整体都画，要么都不画 。 It creates an axis-aligned bounding box that contains all the Meshes, calculates the center point, then uses this information to cull and sort the Mesh instances.

参数特征： 传入矩阵和矩阵的数量 来进行绘制。

限制: You can only draw a maximum of 1023 instances at once.

DrawMeshInstancedIndirect

与DrawMeshInstanced的不同：

• 参数传递方式不同 - ComputeBuffer

• 需要指定bounds

• 没有画多少个的限制

bufferWithArgs： index count per instance instance count start index location base vertex location start instance location

DrawMeshInstancedProcedural

This is similar to Graphics.DrawMeshInstancedIndirect, except that the instance count can be supplied directly using this method.

Benchmark

不同方式绘制的性能

来源：https://www.xuanyusong.com/archives/4488

大量mesh绘制时的性能：

DrawMeshInstancedIndirect（帧率 60） > 自定义Shader中勾选Enable GPU Instancing（帧率 30-40） > 勾选static静态合并批次（40-50帧率）

关于剔除

DrawInstanced接口做的视锥体裁剪，会要么全部剔除，要么全部画出来。而调用之前，所用的矩阵或者是PositionBuffer都可以使用CPU或者ComputeShader算出来，绘制的数量也可以自己决定，因此，可以在绘制之前先做裁剪。