Skip to content

WIP - Support vertex pulling #16826

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 2 commits into
base: master
Choose a base branch
from

Conversation

MiiBond
Copy link
Contributor

@MiiBond MiiBond commented Jun 27, 2025

Vertex pulling is where you do your own reading of vertex data in the vertex shader instead of relying on the standard vertex attribute pipeline. I'm not sure of the best approach in Babylon but I made some minimal changes to make it work. I'd appreciate some feedback and ideas for a better way to set this up.

First, some motivation:
I want vertex pulling as a way of selecting vertex data from neighboring vertices in the IBL Shadows voxelization shader. By retrieving normal info for the provoking vertex of a triangle, I can selectively swizzle the axes of the position to maximize rasterization area and avoid missing voxels. This eliminates the need for 3-pass voxelization and opens the door for doing realtime voxelization of animated geometry (in WebGPU only due to the need for 3D storage textures).
https://playground.babylonjs.com/?snapshot=refs/pull/16826/merge#XSNYAU#124

For vertex pulling to work, we typically need:

  1. An empty vertex layout.
  2. Our needed vertex buffers assigned as storage buffers
  3. A non-indexed draw call with the number of vertices to process. This causes @builtin(vertex_index) to be sequential which is critical for fetching info about neighboring vertices.

No changes are needed to Babylon for the first two requirements. For the 3rd point though, we want to do a engine.drawArraysType even though the mesh has indices. We could mark a mesh as mesh.isUnIndexed = true but then the draw call uses the number of vertices in the geometry and not the number of indices. Another issue is that we don't really want to modify the mesh at all since some render passes won't do vertex pulling. Perhaps a better solution is to do something akin to rt.setMaterialForRendering so that vertex pulling can be enabled per-render pass?
e.g. rt.enableVertexPulling(mesh, true);

Thoughts?

@bjsplat
Copy link
Collaborator

bjsplat commented Jun 27, 2025

Please make sure to label your PR with "bug", "new feature" or "breaking change" label(s).
To prevent this PR from going to the changelog marked it with the "skip changelog" label.

@bjsplat
Copy link
Collaborator

bjsplat commented Jun 27, 2025

@bjsplat
Copy link
Collaborator

bjsplat commented Jun 27, 2025

@bjsplat
Copy link
Collaborator

bjsplat commented Jun 27, 2025

@Popov72
Copy link
Contributor

Popov72 commented Jul 1, 2025

It would be a great feature to have!

However, I think we should strive to improve integration into the engine so that we/the user don't have to manually manage storage buffers and so that it remains as transparent as possible.

Another issue is that we don't really want to modify the mesh at all since some render passes won't do vertex pulling

I wonder if we could just create the vertex buffers with STORAGE flags (in addition to the VERTEX + WRITE flags we currently use)... This way, the buffers should be usable in both pulling and non-pulling modes (to be tested, however, and I don't know if adding unnecessary flags has an impact on performance - STORAGE is not necessary in non-pulling mode, and VERTEX is not necessary in pulling mode).

In this case, we could add a useVertexPulling property to Material. When true, we would not bind vertex buffers to the WebGPU pipeline, but rather bind storage buffers to the shader. We would also inject a “USE_VERTEX_PULLING” define, so that the shader code can adapt accordingly (in case the code wants to support both modes)...

I'm probably missing a few things, but I think this would be a better way to support this mode.

cc @sebavan for the discussion.

@Popov72
Copy link
Contributor

Popov72 commented Jul 1, 2025

Also, in your PG example, I think you would want to calculate the normal yourself by performing the vector product of two edges of the triangle, because the normal to a vertex is generally the average of the normals of the faces to which that vertex belongs, so it is not the true normal to the face.

@sebavan
Copy link
Member

sebavan commented Jul 1, 2025

I really like where it is going :-) and I ll chat with Mike after his break.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants