Skip to content

nvk: Make everything faster

Faith Ekstrand requested to merge gfxstrand/mesa:nvk/push-cbufs into main

This is the perf MR you've all been waiting for. With this MR, we now heuristically promote UBOs to cbufs whenever we can. This means no more load of an address then some math then load of another address. Most uses of UBOs (including crazy update-after-bind cases) now turn into a crazy-fast cbuf src in the instruction.

It's currently marked as Draft because I still have a couple robustness2 corner cases to figure out. However, it seems to mostly be working at this point. App testing (with before/after FPS numbers) would be greatly appreciated.

Merge request reports