RADV Reworking Pipeline Emitting To Improve CPU Usage

[VikingGe]

It's unlikely it actually makes any difference in terms of performance, if that's what you're asking.

It does. I'm noticing this with dxvk at the moment where RADV can quite easily be ~20% faster than AMDVLK when things are CPU bound under some circumstances. And that's massive for a wrapper that spends a vast majority of its execution time in its own state tracker.

[smitty3268]

Bloat can impact performance in many ways, CPU overhead, memory usage, reduced instruction cache residency. Having a whole extra level of wrapping between vulkan app and where the bits go in the command buffer is most definitely not optimal.

Dave.

I'd certainly be more sympathetic to that view if radv was consistently beating amdvlk in benchmarks (of the non-micro variety). Until then, I'd say it's an open question whether the occasional extra cpu overhead really matters since Vulkan uses so much less CPU to begin with.

Edit - I just noticed the reply from VikingGe above about dxvk. If that's true, then I retract my comment here and will join in on the amdvlk bloat bashing.