I recently started tinkering with mesh shaders in my toy engine to get a feel for what is involved in adding support for them and potential performance improvement opportunities and wrote a blog post detailing my findings. interplayoflight.wordpress.com/2025/05/05/m... - ThreadSky

kostasanagnostou.bsky.social • 15 days ago

I recently started tinkering with mesh shaders in my toy engine to get a feel for what is involved in adding support for them and potential performance improvement opportunities and wrote a blog post detailing my findings. https://interplayoflight.wordpress.com/2025/05/05/meshlets-and-mesh-shaders/

Comments

nicebyte.bsky.social•14 days ago

i wonder, how do we account for the fact that there is not PTC involved in mesh shader pipeline? effectively each vert at meshlet boundary is eval'd as many times as there are adjacent meshlets.

nicebyte.bsky.social•14 days ago

hmm, on second thought, as long as caching works within a single meshlet and you optimize the sequence of indices, then cutting up the reordered index buffer into meshlet chunks should probably yield comparable result?

nicebyte.bsky.social•14 days ago

in other words, vcache optimization == optimal meshlet selection. in some sense.

zeux.io•13 days ago

Cutting a vcache optimized sequence into meshlet sized chunks is generally suboptimal, because you get redundant transforms across the split boundary and in practice you get more of them vs an algorithm that is designed to minimize those (latter is meshopt_buildMeshlets, former is buildMeshletsScan)

zeux.io•13 days ago

Re: original question: you do get redundant transforms across the meshlet boundary. The saving grace is that on AMD & NV GPUs, the cache is pretty small and restricted around GPU work scheduling anyway, so meshlets overall end up with fewer transformations as inter-meshlet reuse is sufficient.

pixeljetstream.bsky.social•12 days ago

In that context @nicebyte think of regular vertex shader operating very similar to a mesh shader configured to 32 verts / 32 tris on NV where PD fills the local index buffer, since before mesh shaders existed.

zeux.io•13 days ago

On Intel GPUs, the transform cache used to be ridiculously large but it's been years since I checked the behavior in details; it's possible that on newer, higher end, variants like new desktop cards, the cache has been redesigned to match the larger and more parallel chip as well.

simonmoos.bsky.social•11 days ago

Thanks for the article! In my engine I’ve always created separate vertex buffers for the meshlet data, so I don’t need to do the vertex indirection at runtime. There are plenty of problems with this though.. Have you experimented with the cost of doing the indirection vs doing it directly?

simonmoos.bsky.social•11 days ago

With compute skinning & ray tracing in mind, having the indirection makes life much easier, so I’m probably going to change over to that soon. I’ll try to remember to take some measurements as do!

Comments

Posting Rules

Reply