Cross-Vendor Mesh Shading Being Worked On For OpenGL

[phoronix]

Phoronix: Cross-Vendor Mesh Shading Being Worked On For OpenGL

While it's very rare in recent times for a new OpenGL extension -- especially one that is exciting -- given the continued great adoption of the modern Vulkan API, in 2025 we are looking at an interesting addition to OpenGL with cross-vendor mesh shading via a proposed GL_EXT_mesh_shader implementation...

https://www.phoronix.com/news/OpenGL-GL_EXT_mesh_shader

[edxposed]

Unfortunately, MCRcortex hasn't had time to refactor nvidium lately, so it may remain unavailable to AMD users in the short term.

[Daktyl198]

Unfortunately, MCRcortex hasn't had time to refactor nvidium lately, so it may remain unavailable to AMD users in the short term.

Mesh shaders are the superior way to do it, but in the meantime we've got Distant Horizons which works pretty well. I am perfectly content using that until Nvidium gets refactored (if it ever does).

[aviallon]

Michael the last article (about AUR) lacks a forum entry and redirects to the list of forum entries.

[xcom]

For >RDNA2 right? Not below.

[kiffmet]

Out of interest: How does AMD's mesh shader performance look like these days? AFAIK, they're limited in the sense that a single thread may output one vertex or primitive at most, while Nvidia HW can output multiple. I dunno about Intel though.

[Venemo]

Out of interest: How does AMD's mesh shader performance look like these days? AFAIK, they're limited in the sense that a single thread may output one vertex or primitive at most

They changed that in RDNA3; that limitation is mostly gone now. Details here: https://timur.hu/blog/2024/rdna3-mesh-shading

Nvidia HW can output multiple. I dunno about Intel though.

As far as I know, both NVidia and Intel use a shared-memory-like space where they write all output attributes, so they don't suffer from this limitation. Someone who actually works on that HW can correct me if I'm wrong.

[kiffmet]

Venemo - I missed that blogpost from Timur. Thank you!

As far as I know, both NVidia and Intel use a shared-memory-like space where they write all output attributes, so they don't suffer from this limitation. Someone who actually works on that HW can correct me if I'm wrong.

IIRC Nvidia GPUs can directly address a configurable amount of their L1 cache and use that like AMD GPUs use LDS, but it's faster.

[QwertyChouskie]

Finally, a new SGC post! The chair speaks to us again! 🪑