Tomasz Stachowiak

h3r2tic.bsky.social

14 May 2026 16:37

And, as a convention of ours, any loose resources in a Slang "entry point" module become those strongly-typed params at dispatch site, while param blocks declared in utility modules are supplied by a FrameGraph-like stack. This way you can "push" bulky read-only resources once and they auto-bind.

Tomasz Stachowiak

h3r2tic.bsky.social

14 May 2026 16:37

So that's all fun for scene object draws that need to be as efficient as possible, but in many other cases, you want convenience. The Slang reflection API then enables us to codegen Rust glue for shader entry points so they don't need to rely on stringly matching for one-off dispatches/draws.

Tomasz Stachowiak

h3r2tic.bsky.social

14 May 2026 16:37

Shaders don't need to worry about descriptor sets or bindings then, and the Rust side can cluster those any way it wants. For most of our scene rendering, we use the classic frequency-based grouping, but this way we can easily move bindings around by just changing the clustering on the Rust side.

Tomasz Stachowiak

h3r2tic.bsky.social

14 May 2026 16:37

Instead, in the new backend, our Slang modules use a bunch of parameter blocks which get automatic descriptor set assignment. That creates potentially dozens of descriptor sets tho, so on the Rust side, we cluster those virtual/logical parameter blocks into <= 4 larger ones, and patch the SPIR-V.

Tomasz Stachowiak

h3r2tic.bsky.social

14 May 2026 16:37

Last I checked, Molten could not consume MSL, so we take the Slang -> SPIR-V -> MolvenVK (spirv-cross -> MSL) path. It was mostly about the binding model - Slang's explicit `ParameterBlock` construct is pretty handy for that without having to pre-declare explicit numerical bindings ahead of time.