01:03 <zmike> anholt_: try disabling push descriptors?

01:06 <zmike> there's a thing for it already in init_driver_workarounds on AMD

03:23 <anholt_> zmike: same perf

03:23 <anholt_> (dropping KHR_push_descriptor)

03:23 <zmike> alright, well that's one possibility eliminated

14:15 <zmike> ajax: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18727 if you could

14:16 <zmike> also do you have any thoughts on https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18695

14:41 <ajax> it feels weird to not just pass the full msaa spec into resource_create_front, and have it create both the single-sample and the resolve surfaces

14:41 <zmike> I'm not sure any other driver would care?

14:42 <ajax> d3d12 maybe. enh.

14:42 <zmike> https://gitlab.freedesktop.org/zmike/mesa/-/commit/d042acfc309f7dcb5544caf9dc3a1bba50f49c06 is the part you're considering

14:42 <zmike> or at least how I've done it

14:42 <zmike> https://gitlab.freedesktop.org/zmike/mesa/-/commit/81d4b1c826d5975bd5ba03a8f5d8539266943b59

14:44 <zmike> I think other drivers are more likely not to care since they don't have bespoke wsi handling

14:44 <ajax> tilers too maybe

14:44 <zmike> sure, but I don't think they'll care about the linkage

14:44 <zmike> only that there is a resolve attachment

14:45 <ajax> fair

14:45 <zmike> you can see in https://gitlab.freedesktop.org/zmike/mesa/-/commits/test3/ I've been having a lot of fun with wsi

14:45 <zmike> ~midway down the list

14:47 <ajax> how much of a win is it to push resolves to the swapchain? or are you close enough to it working to measure yet.

14:48 <zmike> it works fine, but it's hard to measure the perf implications since it's related to presentation timing

14:48 <zmike> https://github.com/KhronosGroup/Vulkan-Samples/tree/master/samples/performance/async_compute was my reference

14:49 <zmike> the real issue with it is that it completely explodes validation because layout tracking is broken there (it's correct in zink)

14:49 <zmike> so probably I'd add this as a ZINK_DEBUG to enable it for now

14:50 <ajax> sure but you can turn vsync off and run something slower than 60

14:51 <ajax> anyway. interface a bit wonky but what else is new in wsi.

14:51 <zmike> yep

14:52 <zmike> not sure I have anything that runs at less than 60 fps? unless there's a way to fps limit glxgears or something

14:53 <ajax> surely you have at least one game that isn't full framerate yet

14:53 <zmike> is such a thing even possible?

14:53 <zmike> let me check metro again

14:53 <ajax> for other lesser drivers, i'm told

14:53 <zmike> what even am I looking for

14:55 <ajax> actually, 'glxgears -samples 16 -geometry 1024x1024 -swapinterval 0' you'd hope would show a difference

14:56 <ajax> fairly trivial scene, but still a big resolve hit

14:56 <zmike> hm 16 gives me Error: couldn't get an RGB, Double-buffered, Multisample visual

14:57 <ajax> try 8? depends what glx fbconfigs you have.

14:57 <zmike> that's what I did

15:03 <ajax> maybe just 2 samples. something where the overhead of the decontextualized dri_pipe_blit to resolve takes appreciable time

15:04 <zmike> yeah I'm just testing throughout the branch

15:04 <zmike> looks like it's about a 1-2% improvement

15:04 <ajax> but not where the resolve blit itself takes so much time it dominates

15:04 <zmike> though potentially that's just from having pre-recorded resolve cmdbuf to reuse

15:04 <ajax> you know, still a win

15:04 <zmike> yea

15:05 <zmike> I'll take it

16:19 <zmike> ajax: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18868

16:34 <zmike> Soroush: nice work on those 64bit cleanups