Challenges on Game Rendering

1.Tens of thousands of objects with different type of effects.

2.Deal with architecture of modern computer with complex combination of CPU and GPU

3.Commit a bullet-proof framerate

  -30FPS(60FPS,120FPS+VR)

  -1080p,4K and 8K resolution

4.Limit access to CPU bandwidth and memory footprint.Game logic, network, animation, physics, and AI system are  major consumers of CPU and major memory

Rendering on Game Engine is a heavily optimized software framework to fulfill the critical rendering requiremennts of games on modern hardware(PC,console and mobiles)

 

 

 

 

 

 

 

 

 

 Modern GPU architecture：

 

 Fermi架构

 

 sin cos等特殊运算

 

 冯诺依曼架构，计算和数据分开。数据传输消耗很大，因此尽可能不回读数据,单向传输数据。

 

 

做一次加减乘数a+b可能只用一个clock，但是a在cache中找不到要在内存中找，这要花费一百多个clock。因此尽可能把数据放在一起，在缓存中找到了，cache hit；数据不再缓存中，cache miss。

 

GPU Bounds and Performance 

Application performance is limited by:

-Memory Bounds

-ALU Bounds 数学运算太多了

-TMU(Texture Mapping Unit) Bound

-BW(Bandwidth) Bound

.......

 

 

 

 Renderable

 

 

 

 

 

 smarter way to save data:

 

 

 

 

 

 视觉材质和物理材质是两回事

 

 

 

 

 

 

shader是源码也是数据

 

Render objects in Engine:

 

 

 

 

 

 divided into submesh according to different material.But we will insert all verter into one buffer

 

 

 

 

 

 

 

 rendering每一次为gpu设置了数据，更换数据（set pass call）并不高效,。因此可以把同material的object先绘制，更高效

 

 

尽可能一个drawcall绘制相同submesh和material的instance

Visibilty Culling

视锥外的object其实是看不见的，

 

 

 

 

 

 

 

 要考虑动态物体结构的再构建

 

 

 

 

 

 所谓的开放世界实际也分成了不同的zone

 

 Texture Compression：A must-know for game engine

 

 block-based

 

 

Authoring tools of modeling

Modeling of Polymodeling:3dmax,maya,blender

Modeling of Sculpting:Zbrush

Modeling -Scanning

Modeling -Procedural Modeling:Houdini,Unreal

Clustered-Base Mesh Pipeline:

Divide a complex mesh into a fixed-size cluster(like 32/64 triangle)

 

 

 

 

 

 We can do culling based on mesh clusters instead of object

 

 

In UE5 Nanite:

-Hierarchical LOD clusters with seamless boundary

-Don't need hardware support, but using  a hierarchical cluster culling on the precomputed BVH tree by persistent threads(CS) on GPU instead of task shader

 

Take Away

1.The design of game engine is deeply related to the hardware architecture design

2.A submesh design is used to support a model with multiple materials

3.Use culling algorithms to draw objects as possible

4.As GPU become more powerful, more and more work are moved into GPU, Which is called GPU Driven

 

标签：Engine,into,cache,Rendering,Game,GPU,Modeling,CPU
来源： https://www.cnblogs.com/pisconoob/p/16116521.html