#
#asahi-dev · 2025-04-07 15:57
chaos_princessalyssa: so, before that idea i was typing sth like "just use a llm to rewrite it and check by hand, after all llm output is uncopyrightable", but felt too dirty to post that before you said that
#asahi-dev · 2025-04-07 15:57
alyssahave we considered getting the pdfs from the blockchain and then deep learning them with GPT
#asahi-dev · 2025-04-07 15:56
alyssachaos_princess: 5/10, needs more LLMs
#asahi-dev · 2025-04-07 15:55
chaos_princessby direct link i meant sth like applecdn.com/the-pdf-in-question.pdf that you can hit w/o auth
#asahi-dev · 2025-04-07 15:53
nickchanoh that link does exist https://developer.apple.com/download/apple-silicon-cpu-optimization-guide/ but of course there is a catch
#asahi-dev · 2025-04-07 15:53
chaos_princessright, but unless you get a direct link to apple servers, all of those are license violations, and therefore useless
#asahi-dev · 2025-04-07 15:52
nickchanI only say paste into google for reasons
#asahi-dev · 2025-04-07 15:52
chaos_princessIA? that upload is 100% a license violation,
#asahi-dev · 2025-04-07 15:51
nickchaninto google
#asahi-dev · 2025-04-07 15:51
nickchani mean "Apple Silicon CPU Optimization Guide" -apple.com
#asahi-dev · 2025-04-07 15:50
nickchanwith quotes
#asahi-dev · 2025-04-07 15:50
nickchanthough here is a hack: paste "Apple Silicon CPU Optimization Guide" into Google
#asahi-dev · 2025-04-07 15:49
nickchanand the pdf is /not/ available as a unauthenticated download from apple
#asahi-dev · 2025-04-07 15:48
nickchanchaos_princess: does not really work that way since the json is compiled into the perf bin...
#asahi-dev · 2025-04-07 15:46
nickchanright it's just https://github.com/cyyself/m1-pmu-gen
#asahi-dev · 2025-04-07 15:45
chaos_princessneat
#asahi-dev · 2025-04-07 15:45
nickchanchaos_princess: that already exists
#asahi-dev · 2025-04-07 15:42
chaos_princessmore of a shitpost than a real idea, but what about going for a firmware-style approach, where you have a script that can be fed the pdf and then create the perf jsons?
#asahi-dev · 2025-04-07 15:33
nickchanall short descriptions, less than 20 words, most less than 10 words
#asahi-dev · 2025-04-07 15:25
nickchanlike there's probably only one way (that fits with the kernel style) to write the description of RETIRE_UOP but the same cannot be said for like, FLUSH_RESTART_OTHER_NONSPEC
#asahi-dev · 2025-04-07 15:23
nickchanand there is probably more than one way to write the description for some of the events, that is, possibly involves creativity and therefore copyright
#asahi-dev · 2025-04-07 15:22
nickchandescrption for those events can be found in the apple silicon cpu optimization pdf but of course that's copyrighted
#asahi-dev · 2025-04-07 15:21
nickchancan definitely add the event names at least they are already in the kernel
#asahi-dev · 2025-04-07 15:21
nickchanthere is a bunch of tools/perf/pmu-events/arch/arm64/**/core-imp-def.json in Linux for different cores
#asahi-dev · 2025-04-07 15:20
nickchanso about pmu event descriptions
#asahi-dev · 2025-04-07 07:07
svenif they’re all headed to the entry point you should see m1n1 (or parts of it) and then something might crash because we’re running the entry point again
#asahi-dev · 2025-04-07 07:07
svenmischa85: when you tried to bring up the secondary cores did you see anything on the serial line?
#asahi-dev · 2025-04-06 17:37
svenpity. i was hoping there was a chainload inbetween and the only issue was that m1n1_base changed
#asahi-dev · 2025-04-06 17:37
svenok, that should've worked with the locked iorvbar values we checked earlier
#asahi-dev · 2025-04-06 17:36
mischa85m1n1 on disk
#asahi-dev · 2025-04-06 17:26
svenmischa85: when you tried to start the other cores did you do that with the m1n1 on disk or did you do a chainload to an updated m1n1 inbetween?
#asahi-dev · 2025-04-06 11:36
nickchanand x0 will end up being the return value so check that from the caller
#asahi-dev · 2025-04-06 11:36
nickchanand that function should write a magic value into x0, save x19-x30, wfi, restore x19-x30, then return
#asahi-dev · 2025-04-06 11:34
svenbut first we need to figure out why wfi returns currently
#asahi-dev · 2025-04-06 11:33
sven^-- that
#asahi-dev · 2025-04-06 11:33
nickchanconfigure the timer register to fire an interrupt after a second then call a function that does deep wfi
#asahi-dev · 2025-04-06 11:33
chadmedyeah so we're stuck
#asahi-dev · 2025-04-06 11:32
nickchanand other cores are not working so everything needs to happen on the boot cpu
#asahi-dev · 2025-04-06 11:32
nickchanit keeps the cpu busy
#asahi-dev · 2025-04-06 11:32
nickchanBut that's the opposite of what one wants to do in a wfi experiments
#asahi-dev · 2025-04-06 11:32
chadmedit's exposed to python
#asahi-dev · 2025-04-06 11:32
chadmedmischa85: udelay implements a "timer" by just counting CNTPCT_EL0 ticks so you can use that i suppose
#asahi-dev · 2025-04-06 11:28
nickchanas for polling the timer without causing interrupt that's what udelay() does already
#asahi-dev · 2025-04-06 11:27
nickchanmischa85: you need an interrupt to wake from wfi...
#asahi-dev · 2025-04-06 11:26
mischa85but how can i poll that timer without causing an interrupt? maybe i'm just unsure here how to implement that
#asahi-dev · 2025-04-06 11:24
nickchanand the interrupt of that timer is delivered as a FIQ
#asahi-dev · 2025-04-06 11:24
nickchanthey should be standard ARM stuffs
#asahi-dev · 2025-04-06 11:23
nickchanmischa85: can you not use the cpu register timers?
#asahi-dev · 2025-04-06 11:21
mischa850x302580000
#asahi-dev · 2025-04-06 11:21
mischa85it's defined in the adt tho

plus ›