#asahi-dev · 2022-06-28 11:04
chadmedmps: youll need to change to the schedutil governor btw, wont work with conservative#asahi-dev · 2022-06-28 10:52
chadmedj`ey: CONFIG_ENERGY_MODEL and the schedutil governor is all you should need for it to work#asahi-dev · 2022-06-28 10:47
chadmedmps: you just need the apple cpufreq driver to work. for me it wouldnt load automatically and i had to make a rule for it in modules-load.d so just make sure its actually loaded with lsmod#asahi-dev · 2022-06-28 10:28
chadmedsince the cpufreq driver already registers the energy model if the data is in the device tree, you can just build the DTs from this branch and put them on top of linux-asahi and itll work as long as the module loads#asahi-dev · 2022-06-28 10:26
chadmedhttps://github.com/chadmed/linux/tree/eas-wip#asahi-dev · 2022-06-28 10:26
chadmedwould be nice to get a bigger sample size than 1 though #asahi-dev · 2022-06-28 10:25
chadmedresults so far: slightly cooler running machines, ~2ish hours of extra battery life, no performance regression in day to day tasks, stable#asahi-dev · 2022-06-28 10:25
chadmedif anyone's feeling like doing some testing, i've pushed some devicetree updates for t8103 and t600x that enable EAS with no extra changes. i've been testing myself on a t6000 for ~2 months#asahi-dev · 2022-06-28 07:26
chadmedand the supposed "cluster power" on the firestorm cores goes ballistic whereas the icestorm cores dont seem to be affected at all #asahi-dev · 2022-06-28 07:26
chadmedi added some floating point divisions and FMAs to the "power virus" thing i was working on #asahi-dev · 2022-06-28 07:25
chadmedhm interesting#asahi-dev · 2022-06-23 03:02
chadmedhas anything changed inside m1n1 lately that would now cause iomon.poll() to be throwing exceptions? #asahi-dev · 2022-06-17 00:59
chadmedamarioguy: you shouldnt need to build it in if youre using an initramfs, i dont have anything other than the cpufreq driver and its dependencies built in#asahi-dev · 2022-06-07 01:23
chadmedin fact i didnt touch the code at all#asahi-dev · 2022-06-07 01:23
chadmedbattery life and overall responsiveness seems to be noticeably better with EAS enabled. apple-soc-cpufreq doesnt work as a module, but that doesnt seem to have been anything i did#asahi-dev · 2022-06-05 06:57
chadmedjust update the opp tables and let register_em_with_opp() deal with it (which currently silently fails since we dont have enough data for it in the dt) #asahi-dev · 2022-06-05 06:56
chadmedhaving sane data at this stage at least lets us bring EAS online without having to anything dodgy in cpufreq#asahi-dev · 2022-06-05 06:55
chadmedyeah of course#asahi-dev · 2022-06-05 06:52
chadmedsetting it didnt affect the numbers either#asahi-dev · 2022-06-05 06:52
chadmedi know at least that final set64 is unnecessary, because i didnt copy it and the second cluster still gave me the same values#asahi-dev · 2022-06-05 06:51
chadmedack#asahi-dev · 2022-06-05 06:50
chadmedyeah so im not surprised it would give the same results, just wanted to check#asahi-dev · 2022-06-05 06:49
chadmedthe second p cluster on the m1 max gives the exact same values too#asahi-dev · 2022-06-05 06:48
chadmedeven though the numbers line up with what i had from the SMC...#asahi-dev · 2022-06-05 06:48
chadmedim only unsure because i refuse to believe the entire cluster is only using 200uW on top of the actual core#asahi-dev · 2022-06-05 06:47
chadmedyeah#asahi-dev · 2022-06-05 06:29
chadmedyeah im only really unsure about the p cluster value#asahi-dev · 2022-06-05 06:21
chadmedsurely they wouldnt count in non-reasonable units (tens/hundreds of microjoules, etc)#asahi-dev · 2022-06-05 06:10
chadmedseems true for the firestorm cores at least, im not too sure on the cluster counters or icestorm since i couldnt verify with the SMC but they behave roughly how i'd expect them to (i.e. the values arent totally insane)#asahi-dev · 2022-06-05 06:08
chadmedif i take the difference before and after running bench and divide by the elapsed microseconds, i get values in microwatts that are very very similar to the values i got from the SMC for the pstates where that worked#asahi-dev · 2022-06-05 06:07
chadmedmarcan: i think those regs count up in microjoules#asahi-dev · 2022-06-04 13:35
chadmedcool, ill spend some time on it tomorrow too. i was just playing around with some ideas out while i had some free time before exams so dont feel obligated to go too deep, im sure there are bigger fish to fry :)#asahi-dev · 2022-06-04 13:24
chadmedoh yeah didnt notice the change in set_pstate()#asahi-dev · 2022-06-04 13:20
chadmedarent the clusters at 0x210e00000/0x211e00000?#asahi-dev · 2022-06-04 13:13
chadmedthe one thing that strikes me as odd is the seemingly different units between the pcores and ecores#asahi-dev · 2022-06-04 11:32
chadmedack#asahi-dev · 2022-06-04 11:32
chadmedif you want or have time, if its going to be an effort though dont worry too much about it#asahi-dev · 2022-06-04 11:28
chadmedmarcan: i cant say im shocked either, just trying to think of ways to get values for opp-microwatt without performing invasive surgery on a motherboard#asahi-dev · 2022-06-04 11:26
chadmednew fun challenge: a single core uses such little power in most of the pstates that its basically just noise to the SMC#asahi-dev · 2022-06-04 11:25
chadmedyeah i figured, as i said though turned out the weird errors were due to a dodgy port on the host end, all cleared up now#asahi-dev · 2022-06-04 11:19
chadmedplugged into a thunderbolt port on the actual host machine and all is well now#asahi-dev · 2022-06-04 11:18
chadmedit was the thunderbolt port on my dock, i started getting uart checksum errors too :/#asahi-dev · 2022-06-04 11:05
chadmedi might even just have a dodgy usb cable who knows#asahi-dev · 2022-06-04 11:05
chadmedi did try using the smp_call_sync() instead but it gave me the same results *shrug*#asahi-dev · 2022-06-04 10:32
chadmedamazing what stepping away for a couple of hours and coming back well fed can do :P #asahi-dev · 2022-06-04 10:32
chadmedseemed to have been a race between smp_call and the interpreter iterating over the loop. luckily i seem to be able to get reasonably sane data by just beating the race#asahi-dev · 2022-06-04 08:16
chadmedthats what i expected but the shell just respectfully waits for it to finish doing whatever its doing, which seems to be never because the routine has hung somewhere#asahi-dev · 2022-06-04 08:12
chadmedidk why it would do this since smp_call should always return in time and just run again#asahi-dev · 2022-06-04 08:11
chadmedwhat ive done differently is put the bench routine under a while True, which is what causes the uart timeout#asahi-dev · 2022-06-04 08:11
chadmedim doing that, im following along with what was done to measure the pstate latencies