Daxa 4.0 Ideas

[Ipotrick]

Making incremental changes is becoming difficult and ugly. Some choices were made for the c but especially the c++ api that are not great anymore. A large breaking change could fix a lot of issues with the current api.

Issues/Ideas:

• stl use in public c++ api
  • massive compile time slowdown
  • bad debug performance
• minimize headers, fwrd declare as much as possible
• error handling
  • exceptions are not well liked, cause some minor annoying issues
  • some c api functions return results but the c++ counter part does not
  • unify error handling between c and c++ api
  • improve error type in c++, have a TResult and TError parameters
• remove deprecated api parts
• TaskGraph (maybe this is better suited for a 3.x version?)
  • standalone TaskImage and TaskBuffer are very strange
  • should own resources per default
  • should only optionally wrap an external resource
  • should be tied to some sort of pool or manager (maybe tie them to taskgraphs?)
  • Make TaskAccess's able to be or'ed
• General API
  • remove instance, make it fully hidden. Give each application one static daxa vulkan instance
  • need to rethink bindless image handles, possibly remove all image view api

[GabeRundlett]

threadable SPIR-V compilation utility optional/simpler dependency handling

[GabeRundlett]

Redesign swapchain creation API. Recently #99 was merged which allowed for you to use the correct display transform when writing to the swapchain, but if you want proper HDR support, you still can't select a surface format based on its color space.

I added a terrible "solution" at work, which is to pass color space to the selector. We instead should remove the selector entirely, in favor of a DeviceInfo2 style selection.

That being said, I believe more thought needs to be put into supporting external swapchain images and presentation (maybe external images entirely, but that's a completely different topic). My main use-case for this would be the frame-gen APIs. To my knowledge, FSR3 frame gen needs to handle the swapchain and presentation itself.

[Ipotrick]

add tg managed persistent resources

[Ipotrick]

Tg simplification: only track mip levels, not array layers, move layer tracking to bitfields for simpler slicing

[Ipotrick]

remove puch constant size from pipeline creation, always use 128 bytes

[Ipotrick]

There is too much boilerplate to declare an inline task.

daxa::inl_attachment(daxa::TaskImageAccess::COMPUTE_SHADER_STORAGE_READ_WRITE, task_swapchain_image),

should be much shorter, maybe something like:

daxa::AtComputeShaderStorageRW, task_swapchain_image),

[Ipotrick]

Add daxa::TrasnferMemoryPool docs