#
#asahi-dev · 2025-01-11 13:21
nickchanand T32 instructions has another pmu event
#asahi-dev · 2025-01-11 13:20
nickchanokay here's a problem it turns out that A32 and A64 instructions uses different pmu events
#asahi-dev · 2025-01-11 13:06
nickchanwip code will commit properly later https://github.com/asdfugil/linux-apple/blob/tests/pmu/drivers/perf/apple_m1_cpu_pmu.c
#asahi-dev · 2025-01-11 12:49
nickchanadded A32 EL0 counting support in m1_pmu_configure_counter() but numbers on A7 still seems off (A8-A9) seems normal
#asahi-dev · 2025-01-11 11:56
nickchanand I probably should check out what happens with A32 EL0...
#asahi-dev · 2025-01-11 11:29
nickchan- the available events and affinities are different
#asahi-dev · 2025-01-11 11:29
nickchan- a7 - a9 has 8 counters, a10-m1 has 10 counters
#asahi-dev · 2025-01-11 11:28
nickchan- overall pmu logic is same all the way back to a7
#asahi-dev · 2025-01-11 11:28
nickchanin summary:
#asahi-dev · 2025-01-11 10:24
nickchanit can report instructions and cycles so it works
#asahi-dev · 2025-01-11 10:24
nickchana7 still has 8 counters
#asahi-dev · 2025-01-11 10:11
nickchana10 already has 10 counters
#asahi-dev · 2025-01-11 10:11
nickchana9 has 8 counters only
#asahi-dev · 2025-01-11 09:16
nickchanokay got perf stat to work
#asahi-dev · 2025-01-11 09:09
nickchannevermind I will just redefine m1_pmu_get_event_idx for every table
#asahi-dev · 2025-01-11 09:05
nickchanokay no I can't stash the table in the driver data of cpupmu->plat_device
#asahi-dev · 2025-01-11 08:30
nickchansomehow
#asahi-dev · 2025-01-11 08:30
nickchannow I need a way to test the pmu
#asahi-dev · 2025-01-11 08:09
nickchanokay that worked
#asahi-dev · 2025-01-11 08:04
nickchanokay let's try update mapfile.py
#asahi-dev · 2025-01-11 08:02
nickchangen_driver.py works however
#asahi-dev · 2025-01-11 08:02
nickchanhttps://github.com/cyyself/m1-pmu-gen
#asahi-dev · 2025-01-11 08:02
nickchana13 and below is empty for some reason
#asahi-dev · 2025-01-11 08:02
nickchanwhen i try to generate perf patches with gen_perf_patch.py, only a14 and above worked
#asahi-dev · 2025-01-11 07:10
nickchanI could use the platform_device's data
#asahi-dev · 2025-01-11 07:07
nickchanalternatively somewhere to store a pointer to the affinity table for the pmu
#asahi-dev · 2025-01-11 06:55
nickchanother than trying to comparing the pmu name string
#asahi-dev · 2025-01-11 06:38
nickchanproblem is I need somewhere to store the type of the pmu
#asahi-dev · 2025-01-11 06:35
nickchantrying to see how to refractor m1_pmu_get_event_idx()
#asahi-dev · 2025-01-11 06:19
nickchanwhatever definitely need per-implementation tables now
#asahi-dev · 2025-01-11 06:18
nickchannevermind i messed something up the definitions are very different
#asahi-dev · 2025-01-11 06:13
nickchanhowever some affinities are different
#asahi-dev · 2025-01-11 06:13
nickchanthere's some events on m1 that's not on a7 (expected)
#asahi-dev · 2025-01-11 06:12
nickchanUgh
#asahi-dev · 2025-01-11 06:10
nickchani should generate A7 definitions first and see how different it is from m1
#asahi-dev · 2025-01-11 06:10
nickchanno
#asahi-dev · 2025-01-11 06:07
nickchanokay i don't really know what I am doing but i am going to add a11 pmu into the dts and see whatever happens
#asahi-dev · 2025-01-11 05:54
nicolas17nickchan: btw if you ever feel like writing docs, I think more technical hardware info is totally in scope for theapplewiki ;)
#asahi-dev · 2025-01-11 05:49
nickchanamazing macOS has pmu definitions for Apple A7
#asahi-dev · 2025-01-11 05:37
nickchanonly l1 and l2 listed
#asahi-dev · 2025-01-11 05:37
nickchanany particularly reason why system level cache is not listed in dts file for M1?
#asahi-dev · 2025-01-10 13:10
leioand I'm a snowflake that would WANT to do that even on 4k capable hw for the perf improvement
#asahi-dev · 2025-01-10 12:08
chadmedwe're special snowflakes in the sense that we really _cant_ do that
#asahi-dev · 2025-01-10 12:08
chaos_princessyep
#asahi-dev · 2025-01-10 12:07
chadmedyeah but distros will simply ship 4k kernels on those machines anyway
#asahi-dev · 2025-01-10 12:07
chaos_princesslike, msft is pushing WoA hard
#asahi-dev · 2025-01-10 12:07
chaos_princessnot really?
#asahi-dev · 2025-01-10 12:06
chadmedi think our use case is pretty unique tbh. other arm hw consumers are not likely to be requiring such heavy x86 emulation
#asahi-dev · 2025-01-10 12:06
chaos_princessthe only other arm board i know of that also cares about page sizes is the big nvidia chips
#asahi-dev · 2025-01-10 12:05
chadmedif it works in muvm on other hw, the issue is probably hk etc etc

plus ›