#asahi · 2021-09-19 01:57
marcanit's a pretty name everywhere except in spain :p#asahi · 2021-09-18 18:11
marcan(that's already enabled on mine)#asahi · 2021-09-18 18:10
marcanI bet it's a firmware difference then#asahi · 2021-09-18 17:56
marcanoff to sleep :)#asahi · 2021-09-18 17:55
marcanthen linux doesn't have to care#asahi · 2021-09-18 17:55
marcanif we end up needing init, I'm tempted to just do it in m1n1 (and also default the pcores to at least 2GHz)#asahi · 2021-09-18 17:49
marcanso better use 1#asahi · 2021-09-18 17:49
marcan0, but that's 24MHz which is slow enough it breaks the setup.py process#asahi · 2021-09-18 17:48
marcanI'm going to *guess* either the LIMIT things or the CLUSTER_DVMR thing#asahi · 2021-09-18 17:48
marcanj_ey: can you bisect what did it?#asahi · 2021-09-18 17:47
marcanha, so either it's an Air thing or a firmware thing#asahi · 2021-09-18 17:47
marcananyway, I should get some sleep :)#asahi · 2021-09-18 17:45
marcanthis is the actual waste-of-time-on-stream poke all the things version, if you want to try it too: https://mrcn.st/p/L92lzFKJ#asahi · 2021-09-18 17:44
marcanah yes#asahi · 2021-09-18 17:38
marcandcp.py is all tested on 12.0 beta5 (I should update to beta6...)#asahi · 2021-09-18 17:37
marcan(that was kind of the point of going to 12.x, getting ahead on those ABI changes, which is why I don't want to support 11.x at all for the initial merge of this)#asahi · 2021-09-18 17:37
marcanI thought you were on 12.x, with the 4 slots for surfaces and such#asahi · 2021-09-18 17:36
marcanalyssa: so you're doing the DCP driver on the 11.x ABI?#asahi · 2021-09-18 17:29
marcanI was using cpu7 in parallel to test the boost thing, forgot to remove it#asahi · 2021-09-18 17:29
marcanj_ey: fixed#asahi · 2021-09-18 17:27
marcanactually, that was a hack#asahi · 2021-09-18 17:27
marcanj_ey: no#asahi · 2021-09-18 17:27
marcanI assume you're not on 12.0 like I am?#asahi · 2021-09-18 17:27
marcanalso, this could be a firmware difference#asahi · 2021-09-18 17:25
marcanj_ey: use cpu_pstates.py to just test the highest states. there's a pile of commented out stuff there. see if uncommenting any/all of it fixes it.#asahi · 2021-09-18 17:24
marcanthere might be some voodoo needed to get things working on the air, if this isn't just a throttling thing#asahi · 2021-09-18 17:23
marcan(briefly)#asahi · 2021-09-18 17:23
marcanalyssa: fwiw it will do 3.2 on battery on macos#asahi · 2021-09-18 17:22
marcaneither that or we actually do need more init on the air#asahi · 2021-09-18 17:22
marcanpretty sure you're throttling#asahi · 2021-09-18 17:19
marcanlol#asahi · 2021-09-18 17:19
marcandifferently?#asahi · 2021-09-18 17:18
marcanbad command is one obvious signal for an ABI break ;)#asahi · 2021-09-18 17:18
marcanthere have been incompatible changes#asahi · 2021-09-18 17:17
marcanj_ey: did you chainload?#asahi · 2021-09-18 17:04
marcanI would love to see how fast it runs llvmpipe with all that :p#asahi · 2021-09-18 17:04
marcanalyssa: added experiments/cpu_pstates.py, that will do the memory config too#asahi · 2021-09-18 16:58
marcanalso note that that set64 is missing the memory controller performance bump, which should increase memory bandwidth on top of that#asahi · 2021-09-18 16:57
marcan*big.LITTLE#asahi · 2021-09-18 16:57
marcanalyssa: also you probably want to taskset any benchmarks to cpus 4-7, since linux has no idea about bit.LITTLE with our current devicetrees#asahi · 2021-09-18 16:56
marcanbut this is good enough to get the ballpark right, which is all that matters for this#asahi · 2021-09-18 16:55
marcanthe timer is 24MHz, so there's quite a bit of rounding/jitter error#asahi · 2021-09-18 16:55
marcane.g. the pipeline blip number really has ~1 significant digit#asahi · 2021-09-18 16:55
marcannote that these numbers have, er, quite some error margin#asahi · 2021-09-18 16:54
marcanI think intels stumble for a few microseconds#asahi · 2021-09-18 16:54
marcanalso the maximum actual pipeline blip during these transitions is only ~350ns, which is very nice#asahi · 2021-09-18 16:53
marcanthis makes sense since upclocking has to increase voltage#asahi · 2021-09-18 16:53
marcanwhile upclocking takes 20~55µs depending on how many steps up you're going#asahi · 2021-09-18 16:53
marcanso basically downclocking takes a fixed ~6µs latency across the board, very roughly#asahi · 2021-09-18 16:52
marcanj_ey: yeah