Question Skinning in Vertex Shader vs Compute Shader?

Which would be more efficient: character skinning directly in the Vertex shaders and streaming out the skinned data for reuse in shadows or just using a Compute Shader?

I believe Compute shader path to be faster since its newer than Stream-Out but since I will be drawing hundreds of skinned characters together using instancing I guess I will need an extra buffer to store the skinned vertices for each instance. Is there a smarter way?

 

Compute is more ideal since you're not invoking the whole raster pipeline just to update a buffer. And makes it much easier if you want to add other shader passes since the data will simply exist already in its modified form.

You can access the Mesh's vertex buffers directly on the GPU, so you don't need to double-buffer it.

https://github.com/Unity-Technologies/MeshApiExamples

Just modify the mesh buffer directly and now the meshes will automatically be rendering with those updated values.

 

Hmmm I know I can modify the vertex Buffers directly, but won't there be some problems over time due to precision issues? Otherwise you're right it will be faster to just dispatch a compute shader and directly modify. Btw after studying more I realize that stream-out is a somewhat obsolete method after DX11 Compute shaders and RWBuffers. So I guess I won't be using it

 

If you're continuously changing the shape there could be some divergence in shape overtime yes, in which case you would want to either double buffer it, or simply store the original vertex positions into an unused UV channel (or vertex colors) which your compute shader can then use as the starting point each time you modify.

 

That is a good idea never thought of storing in the UVs, but since I need to store Normals and Tangents too I am thinking of storing them in a separate buffer, mesh data don't take that much memory anyway and I don't have that many characters.

 

Well, you do have 8 UV channels of which you're likely only using 1 or 2 for character stuff

But yeah whichever you feel like!

 

Unity's SkinnedMeshRenderer uses separate streams, and then only binds the relevant ones, to improve vertex cache usage. A skinned mesh has essentially 5 buffers of vertex data...

A: Original vertex position/normal/tangent
B: Transformed position/normal/tangent
C: Previous frame position/normal/tangent (for motion vectors and stuff)
D: Bone weights and indices
E: UVs (and vertex colors if it needs them)

Then it binds what it needs for each pass. So A and D are only bound for the skinning stage and not used during rendering.

 

Oh right I completely blanked they said character... in which case yeah just use Unity's own pre-skinning buffer haha