11:20 <fredldotme> Hello! I'm experimenting with changes in the display stack on Ubuntu Touch on the PinePhone right now, and I got to the point where both old Mirclient and new Wayland clients work fine! I've run into performance issues though. Is my assumption correct that it's the creation of KMS dumb buffers for every frame + triple buffering is causing the performance drop?

11:21 <fredldotme> In other words: I had to change Lima to get me a scanout whenever a PIPE_BIND_SHARED gbm_bo is requested, and that seems inefficient

15:59 <anarsoul> fredldotme: BO allocation is expensive, don't do that for every frame

16:00 <fredldotme> Thanks for the answer anarsoul, what would your suggestion be?

16:00 <fredldotme> Just reusing BOs across buffer passings?

16:02 <fredldotme> Little background: We're trying to support old Mir apps (integrated into the shell) and Wayland apps (the future of course) until we can Wayland-ify everything. Seems though that the BO allocations are less expensive on Virgl, Intel etc than Lima

16:12 <anarsoul> fredldotme: reuse BOs, make sure that draw that was using the BO is complete before reusing it

16:12 <anarsoul> fredldotme: it's expensive there as well (that's why all the drivers have BO cache)

16:12 <anarsoul> probably you don't notice it since CPU is more powerful

16:12 <anarsoul> anyway, use perf to profile your app to find the bottleneck

16:16 <fredldotme> Thanks, will try

20:07 <anarsoul> enunes: rellla: should we switch to genxml for generating some binary structures like texture descriptor, RSW, possibly GP and PP instructions?

20:07 <anarsoul> at least for TD and RSW currently it's a mess...