Gokhan Avkarogullari

This may not be exactly what you want, but I don't think MS or IHVs wanted to have to support multiple D3D driver stacks indefinitely. D3D11 drivers are expensive if you want them to be fast, emulation of the higher-level API on the lower one is probably the best choice.

IMO nobody knows "best performance practices" for given hw generation as well as IHVs, they should be the ones mapping some intermediate or high level constructs to whatever hardware consumes (with opt-in super low level hints, constructs, extensions)

Dx12 allows devs to expose information only the ISV knows. We understand what memory is needed when, what materials we’re about to draw, what the frame is about to do. We can argue that a simpler API is possible, but trying to optimize/guess this in drivers WAS the problem

I think there is a disconnect between our experiences as I *never* encountered any serious optimization that was so specific to my use-case vs API as opposed to Nvidia vs AMD (like whole handling of constant buffers vs constant fetches)

Multithreaded submission and guaranteed shader compilation don't count? I mean, those are two of the big ones. Stall to compile when you first draw was an API issue, not an IHV issue.

Oh this is different - definitive flaw in DX11, not dependent on the application, could be fixed in 11.3 or .5. I'm asking about something that only software developer knows and is so unique no reasonable driver can handle.

hmmmm "I need these shaders compiled, so here is all the state I will use that might affect the bytecode" seems like it would qualify? Sure d3d11 could have been made incrementally better, but switch to PSOs isn't an incremental change. That would be a rewrite.