Skip to content

radv: Refactor/rewrite acceleration structures

This MR aims to convert all the acceleration structure build kernels to GLSL/C++. The benefits will be:

  • It will be easier to implement non trivial kernels.
  • We can share code between the CPU and GPU path.
    • This will also enable us to work on new algorithms without ever using the GPU. (CPU builds + some form of BVH validator/debugger)

This is done in 3 parts:

  1. Split radv_acceleration_structures.c into a common, CPU and GPU file.
  2. Port all the existing kernels to GLSL.
  3. Add CPU 'dispatch' code.

Why C++?

C++ is more compatible with GLSL than C because

  1. We don't have to always add the struct keyword.
  2. We can use operator overloading to implement the builtin GLSL types.

TODOs:

  • Port the internal nodes kernel.
  • Port the copy/[de]serialization kernel. I don't know how the serialization code works and I won't know it anytime soon.
  • Port the leaf nodes kernel.
  • Add the dispatch code for CPU builds.
  • Compatibility with older generations. (Without atomic float min/max)
  • Run CTS. (nothing broke)

Just opening this MR to get some early feedback about the general idea :)

/cc @bnieuwenhuizen @allanmac

Edited by Konstantin Seurer

Merge request reports