bevy_pbr: Reuse the mesh.wgsl vertex stage for shadow mapping and remove depth.wgsl

[superdump]

Objective

• Reduce fragmentation of vertex stage code which can easily lead to bugs and mistakes as depth.wgsl is 'missed'

Solution

• Refactor the View struct out of mesh_view_bind_group.wgsl into view_struct.wgsl and define the bevy_pbr::view_struct import. Import this into mesh_view_bind_group.wgsl as it is used there for bindings.
  • NOTE: The types, bindings, and functions are also refactored into separate files for easier reuse in Reusable PBR shader types/bindings/functions #3969
• Remove depth.wgsl and rework mesh.wgsl to reproduce the same vertex stage by excluding unneeded code with the help of a SHADOW_MAPPING shader def

---

Changelog

• Changed: Shadow mapping passes now use the same vertex shader file as main passes to ease maintenance

Migration Guide

• Use MESH_SHADER_HANDLE with the SHADOW_MAPPING shader def for the vertex stage of shadow mapping passes instead of SHADOW_SHADER_HANDLE

[james7132]

For this to be correct in the general case, we need to inherit the shaderdefs of the specialized mesh pipeline in addition to this one.

[superdump]

The point was that with SHADOW_MAPPING defined, only code that is either without any other shader defs or that has shader defs that are configured for shadow passes are included.

[superdump]

That is a ‘by convention’ maintenance burden, but I think it’s pretty manageable because so little is needed for shadow mapping. Just clip position.

[superdump]

I guess your comment was about, for example, VERTEX_TANGENTS. My point is that if SHADOW_MAPPING is set, no other shader defs are part of what remains.

[james7132]

This is likely going to be needed when we merge in skinning and morph target support. Where they are needed in the shadow vertex shader to be correct. Any other user made specialized define would also likely need to be reflected here too.

[superdump]

Yeah, I realised what you meant when I looked over the mesh skinning PR last night. I'll have to figure out how to hook that up then.

[superdump]

I'm going to close this because I want to move in the opposite direction to a more specialised depth pass. The problem with reusing the mesh vertex stage is that it needs to load all vertex attributes. For non-animated geometry for depth-only passes, only the position attribute is needed. I have learned since making this PR that it is a good idea to have separate vertex buffers for the position attribute and for everything else so that depth passes do not even read the unused vertex attribute data, have better cache usage, and improve performance.