Even with latest wine and the dispatcher optimizations, this MR still regresses [vkoverhead](https://github.com/zmike/vkoverhead) testcases significantly, for example: - test 46 updating a single UBO with vkUpdateDescriptorSets from 48000 to 42000, ~12,5% regression - test 92 using descriptor_buffer from 84000 (with !1552 for direct calls) to 62000, ~26% regression
I'm sorry to ask for justification for a change as an outsider from the project, but that does bring up the question of whether getting rid of direct calls is as urgent as to justify such a large regression. Are there any plans to keep direct calls around for longer in Proton? Especially considering Proton's target platform, which has a relatively weak, battery powered CPU, I don't see how even adding 5% overhead to very hot vulkan functions would be acceptable.