01:03joebobjoe: how do I access the hardware planes of nvidia gpus? does nvidia provide any documentation on them?
01:22joebobjoe: wait, do gpus not have hardware overlays anymore? what?
01:35joebobjoe: hm. I guess what I am thinking of is called hardware compositing
01:35joebobjoe: didn't know that was separate from overlays
20:04RSpliet: karolherbst: mind giving this branch a spin with a few benchmarks? https://github.com/RSpliet/mesa/tree/insn_sched
20:05RSpliet: On my Kepler: Unigine Heaven low settings 30.3->32.2fps, Xonotic 75->76fps
20:05Arcanos: Just an update, The Nvidia QUadro FX380 that I have, Works Wonderfully
20:08RSpliet: oh, and pixmark piano went from (133-140)ms per frame to (123-130)ms
20:29karolherbst: RSpliet: sounds good :)
20:30RSpliet: karolherbst: Thanks. 'd like to hear if you can observe similar gains (or gains in different workloads)
20:31RSpliet: Before I seek dboyan to see if he's happy for me to use his work this way, and imirkin to provide feedback
21:28RSpliet: gosh, I've been so busy all this time that I never realised how easy it is to use my laptop's GM107 at full speed
21:28RSpliet: let's re-run some benchmarks on there
21:50karolherbst: RSpliet: :D
23:10RSpliet: Same strategy on GM107 gives only 0.3fps more in heaven... that's within the "negligible" margin. I guess the combination of Maxwell and "twice the DDR3 at over 2GHz (SDR clock? Can DDR3 even do that?)" might just shift the bottlenecks or requires even more cleverness
23:12RSpliet: also, heaven randomly segfaults on GM107, with and without the new pass. Doesn't do that on GK107
23:16RSpliet: Wait, how does this even... it's a "940M" GM107 (NV117)... both wikipedia and our codenames list claim I should have a GM108.
23:26RSpliet: Oh... so the Maxwell wiki page lists both. And there's a monumental difference between "940M" GM107 and GM108. Like, the former (mine :-D) has ~double the DRAM bus width, shader cores, TDP and single-precision perf. Marketing names...