#asahi-dev · 2023-03-11 12:56
chadmedapple-pmgr-pstate 28e080000.power-management:power-controller@2XX: failed to add to parent domain: -22#asahi-dev · 2023-03-11 12:43
chadmedlet me try again hang on im in macos #asahi-dev · 2023-03-11 12:43
chadmedyeah new kernel, new initramfs, ran update-m1n1 pointing at the correct dts#asahi-dev · 2023-03-11 12:39
chadmedit fails to boot, cant find the power domain for (seems to be) nvme #asahi-dev · 2023-03-11 12:39
chadmedno not that one #asahi-dev · 2023-03-11 12:38
chadmedasahi-6.2-5 has that pmgr failure on j314 here. dts are correct and everything built fine#asahi-dev · 2023-03-11 12:30
chadmedlatest u-boot + sd-boot works fine on j314 here too#asahi-dev · 2023-03-04 15:00
chadmedradeon first? the one where every patch is like 400,000 lines of autogenerated register definition headers? #asahi-dev · 2023-03-04 14:57
chadmedi gave up trying to read it very very quickly so yes#asahi-dev · 2023-03-04 14:51
chadmedi thought as much :P#asahi-dev · 2023-03-04 13:56
chadmedeiln: i mean thats just begging for an MR adding all that stuff in after reverse engineering it#asahi-dev · 2023-03-04 13:39
chadmedcache and gddr phys yeah, mb#asahi-dev · 2023-03-04 13:37
chadmedthough the infinity fabric phys are pretty big themselves#asahi-dev · 2023-03-04 13:36
chadmedmaybe we can call navi 31 a true gpu since (most of) the non-shader core stuff got offloaded to the chiplets#asahi-dev · 2023-03-04 13:30
chadmediirc nvidia cards have two encoding blocks too #asahi-dev · 2023-03-04 13:28
chadmedespecially now as vendors are starting to tack on AV1 support to all these blocks#asahi-dev · 2023-03-04 13:27
chadmedoh that happened, neat#asahi-dev · 2023-03-04 13:26
chadmedneural engine#asahi-dev · 2023-03-04 13:26
chadmedwhen eiln is done there will be three apple drm drivers! #asahi-dev · 2023-03-04 13:26
chadmedwho cares who makes them #asahi-dev · 2023-03-04 13:26
chadmedyeah it doesnt really matter, they are functionally distinct pieces of hardware#asahi-dev · 2023-03-04 13:23
chadmedthey always say "GPU" to mean the entire card#asahi-dev · 2023-03-04 13:23
chadmedi havent heard anyone say "graphics card" in a long time#asahi-dev · 2023-03-04 13:22
chadmedit was some time in the 2010s though#asahi-dev · 2023-03-04 13:22
chadmedin fact i dont even remember when we stopped calling them "graphics cards" and started calling the whole thing the GPU#asahi-dev · 2023-03-04 13:22
chadmedi feel like drilling this distinction into peoples' heads is going to be the hardest part of any hypothetical transition away from the IBM-compatible#asahi-dev · 2023-03-04 13:16
chadmed(also apple was taken by the dcp driver) #asahi-dev · 2023-03-04 13:16
chadmedand alyssa called the mesa driver asahi so it makes sense to keep the kernel driver nomenclature the same#asahi-dev · 2023-03-04 13:16
chadmeddifferent drivers for different hardware blocks#asahi-dev · 2023-03-04 12:02
chadmedthe interfaces exist and we know they work (see https://gist.github.com/chadmed/e889ea905876db29cbebb7265039f5a8) so we should be using them#asahi-dev · 2023-03-04 12:00
chadmedbut basically the only way to fix this is for userspace to start communicating its needs to the scheduler in a more precise way than not at all#asahi-dev · 2023-03-04 11:55
chadmedwell, i say "very" in the sense that its noticeably slower but its still snappier than any other machine ive owned#asahi-dev · 2023-03-04 11:53
chadmedthe detrimental effect of EAS is very noticeable on m1 pro since the two ecores just end up crammed with tasks #asahi-dev · 2023-03-04 11:52
chadmedid like to see what kind of effect EAS has on interactivity on socs with 4 ecores, since theres double the ecore capacity#asahi-dev · 2023-03-04 11:51
chadmedso the scheduler goes "eh good enough" and throws everything at them#asahi-dev · 2023-03-04 11:51
chadmedtypically you see ecores at ~half the performance of the pcores but we retain ~70% of the performance#asahi-dev · 2023-03-04 11:51
chadmedi also think we're an interesting case in that our cores arent _that_ asymmetrical in terms of performace but are in terms of power consumption#asahi-dev · 2023-03-04 11:42
chadmedmaybe we should talk to the kde eco folks, this seems like something theyd be interested in#asahi-dev · 2023-03-04 11:41
chadmedbut one i think is absolutely unacceptable when we have mitigation tools that are yet to be exhausted#asahi-dev · 2023-03-04 11:41
chadmedamd are brute forcing their way around this on windows with a driver that just makes non-vcache clusters unavailable when it thinks youre running cache-heavy loads and vice versa which is... an interesting approach#asahi-dev · 2023-03-04 11:39
chadmedre schedutil interactivity i dont think there are any real gains to be had anymore just by tweaking CFS and schedutil. the last "big" thing was arm working on util_est to help offset the effects of util decay#asahi-dev · 2023-03-04 04:20
chadmedis that something youd be able to catch by writing a metal shader that _should_ use a cmpxchg and then decompiling it to see what it's doing?#asahi-dev · 2023-03-04 02:02
chadmedhell mine goes brrrr enough to cause coil whine during some glmark2 tests#asahi-dev · 2023-03-04 02:02
chadmedas long as it goes brrrrr i think people will be happy#asahi-dev · 2023-03-04 02:00
chadmedmaybe glide would be a good fit!#asahi-dev · 2023-03-04 01:59
chadmedor is it much of a muchness#asahi-dev · 2023-03-04 01:58
chadmedand aiui vulkan should be a better map to the hardware anyway (geometry and tessellation notwithstanding)? #asahi-dev · 2023-03-04 01:57
chadmedyeah that makes sense, especially with zink#asahi-dev · 2023-03-04 01:55
chadmed"oooh we dropped 3 cycles in the culling pipeline" okay but tetris is a pink rectangle #asahi-dev · 2023-03-04 01:54
chadmedand theres no point worrying about fastidious optimisation while theres stuff that just like... doesnt work at all