Animated Crowd Rendering

From GPU Gems 3by Bryan Dudash

Johan WallermarkJonas E. Mattsson

10 000 is a crowd

DX9 vs DX10

DX9Hard to achieve instancingUnable to index into constant memory based on current instanceCache-inefficient work-around exists

DX10Instancing is at the core of the APINew functions to support drawing multiple meshesVarious system variables => generated by the GPU

Instancing

Render a mesh multiple times with different locationcolor, etc

Usually static meshese.g. grass, leaves, other things that one needs a lot ofcan't be in different animations

Uses constant memory to store dataUsed to transfer workload from CPU to GPU Skinned Instancing

Same as above but uses texture memory to store data

Skinned instancing

How does this work?

SV_InstanceIDTexture stores all animations

texture used instead of constant memory because of the amount of dataoffset used to access the right animationtexture fetches are fasttexture animation is a multiple of 4

easy to get the next matrix by just offsetting the u component

Ze math

4096 float4 maximum under DX10The papers example uses 5 float4 for their data=> 4096/5 ~= 819 instances per draw callTo draw 10 000 characters=> 10 000 / 819 ~= 13 draw calls

13 draw calls is much better than 10 000 (if only one mesh)

Rendering flowCPU

Perform game logic (animation time, AI, etc.). Determine an LOD group for each instance and populate LOD lists.For each LOD:

For each submesh:Populate instance data buffers for each instanced draw call.

For each buffer: Call DrawInstanced() on the submesh.

Rendering flow cont.

GPUVertex Shader1. Load per-instance data from constants using SV_InstanceID.2. Load bone matrix of appropriate animation and frame.3. Perform palette skinning.

Pixel Shader1. Apply per-instance coloration as passed down from vertex shader.2. (Optional) Read a custom texture per instance from a texture

array.

Geometry variations

Having all characters look the same with the same animation looks really weird and humans will by a first glance see that something is off.hence breaking characters into several different pieces and instancing those separately

Performance increasers

Level of detail - LODsets LOD groups on the CPU

View frustum culling (CPU)

Performance

Intel Core 2 Duo 2.93 GHz, 2GB RAM, NVIDIA 8800 GTX @1280x1024 9547 characters34 frames/secpixel/fill limited LOD 20, 60 & 100 unit radiuses160 instanced draw callsvs 59 726

Own comments

Shadows for 10 000 characters, hmm...Polygons per character Various effectsPerformance

59 726What method?

Thank you for your time

Questions?