Philpax

https://fosdem.org/2026/schedule/event/QU9UF3-slang-cross-platform-inference/

• Slang
  • originally a NVIDIA initiative, but donated to Khronos in 2024
  • in a way, superset of WGSL/WESL/Naga
  • programming language for the GPU, but also the compiler for that language
• So why Slang?
  • Open-source, open-governance
  • Write your shader once, runs everywhere
  • Supports both compute and rendering
  • Not limited by the lowest-common-denominator, like WebGPU
  • Autodiff built in, but Crozet hasn't used it; could be useful for traning
  • Most importantly, it has modern syntax; so much nicer than WGSL or GLSL
  • Supports reflection; can ask the compiler for information about your shaders. Can have the compiler allocate binding sets, and then ask the compiler how it was done
  • See Khronos's BOF presentation at SIGGRAPH 2025: https://www.khronos.org/assets/uploads/developers/presentations/Slang_BOF_SIGGRAPH_2025.pdf
• Using Slang in Rust
  • The compiler is written in C++, but there are Rust bindings
  • Crozet wrote a small wrapper (minislang) to make it nicer to work with
  • Need to write a RHI (rendering hardware interface?) to translate Slang's commands to each graphics API - slang-hal - only for the GPU API, not for the shaders
  • Utilities/libraries built atop slang:
    • stensor for tensor manipulation - new library because he needs these operations to be reusable from other domains
    • slai - physics simulations etc
• Slang in action
  • Structured buffers are not global variables, like they are in other shading languages, but actual arguments to kernels; makes it easier to maintain
  • Slang has a built-in module system, so you don't need a preprocessor or anything like that; can reuse code across your codebase or other libraries
  • Write kernel launch - no binding groups, no backend choice - it's all handled for you
• slai: dimforge is working on Slang AI
  • Very incomplete right now, still in the experimentation phase
  • Have implemented Gemma + Whisper model
  • See wgml for a purely WebGPU implementation
• What's next?
  • Still targeting the single-source GPU scientific computing dream
  • Close to ideal, but no package manager, no CPU/GPU code sharing, incomplete IDE support
  • Last experiment; Rust-GPU, more at the end of the year
  • Also have a focus on generating datasets for AI stuff using GPU-based phyics simulations
• Question: Does it have support for generics?
  • [wasn't paying attention]
• Question: Game engine integration?
  • Theoretically possible, just need to be careful about staying in the same API
  • Bevy is exclusively WebGPU; Slang could generate the WGSL, and use them directly