#asahi-dev · 2026-06-03 21:39
yukaBut not the second pcore#asahi-dev · 2026-06-03 21:39
yukaI can measure the first pcore (boot core) of the Neo at the highest expected clock#asahi-dev · 2026-06-03 19:43
yukabut we are not allowed to do that :(#asahi-dev · 2026-06-03 19:43
yukaReading xnu code, it should configure sleep through ACC_OVRD like on other devices#asahi-dev · 2026-06-03 19:42
yukaBecause of the cpufreq things, I started looking into sleep behavior on M4 again#asahi-dev · 2026-06-03 18:40
yukaI'm also seeing it on A18 Pro and M4 (even with switching to an e-core), probably because they don't sleep properly#asahi-dev · 2026-06-03 18:40
yukayep#asahi-dev · 2026-06-03 18:27
yukajannau: only if I switch the boot core to a core in another cluster, see the other commits in https://github.com/yuyuyureka/m1n1/commits/pstate-latencies-boot-cpu/#asahi-dev · 2026-06-03 18:21
yukanow better?#asahi-dev · 2026-06-03 18:20
yukaAh well#asahi-dev · 2026-06-03 11:13
yukaIs the firmware interface even that different from previous ones?#asahi-dev · 2026-06-02 21:35
yukaand *t6030#asahi-dev · 2026-06-02 21:35
yuka*measure pstate latencies#asahi-dev · 2026-06-02 21:35
yukajannau if you want to measure clocks on t8122 or t6031, try this branch: https://github.com/yuyuyureka/m1n1/commits/pstate-latencies-boot-cpu/ (you need the cregs and number of pstates but the rest should work now)#asahi-dev · 2026-06-02 17:13
yukaI guess it's a _target_ p-state#asahi-dev · 2026-06-02 17:11
yukaI heard about those p-states which can only be reached if some other cores are in deep wfi. Does Linux somehow know about this for scheduling, or will it still attempt to use the highest p-state even in an all-core workload, and what will happen then?#asahi-dev · 2026-06-02 16:53
yukaso I should measure the e-core cluster from p-core, and the p-core cluster from e-core#asahi-dev · 2026-06-02 16:53
yukaHmm, right#asahi-dev · 2026-06-02 16:53
yukathe other cores shut down completely?#asahi-dev · 2026-06-02 16:52
yukathe only core how?#asahi-dev · 2026-06-02 16:51
yukaodd indeed that the frequency goes down again#asahi-dev · 2026-06-02 16:50
yukajannau: no, the e-core pstates match with powermetrics output#asahi-dev · 2026-06-02 16:48
yukahttps://yuyu.yuka.dev/public/t6034-powermetrics.txt#asahi-dev · 2026-06-02 16:43
yukaThen it is already in the right place#asahi-dev · 2026-06-02 16:42
yukathe header says t0631-base is for t603[124] but not t6030#asahi-dev · 2026-06-02 16:42
yukaAh no#asahi-dev · 2026-06-02 16:41
yukaSo this table needs to go in t6031.dtsi instead of t6031-base.dtsi, correct?#asahi-dev · 2026-06-02 16:40
yukachaos_princess measured similar things on t6031 using cpu_pstate_latencies.py#asahi-dev · 2026-06-02 16:39
yukaT6034 powermetrics match my table (except of higher p-states missing in my table)#asahi-dev · 2026-06-02 16:15
yukaCrazy that this thing can go even faster, it already feels super fast#asahi-dev · 2026-06-02 16:14
yukaAhh okay then I need to re-do the table#asahi-dev · 2026-06-02 15:57
yuka(maybe on purpose, so that I focus on the other things I should be doing ;))#asahi-dev · 2026-06-02 15:57
yukadon't have a second device or central scrutinizer with me rn#asahi-dev · 2026-06-02 15:56
yukaI will check it later#asahi-dev · 2026-06-02 15:54
yukaDo you maybe have less (actually distinct) p-states on your M3 Pro?#asahi-dev · 2026-06-02 15:54
yukaso then these could have different p-states?#asahi-dev · 2026-06-02 15:54
yukaIt lists M3 Max but not M3 Pro#asahi-dev · 2026-06-02 15:43
yukaor was that only for the 16-inch max previously...#asahi-dev · 2026-06-02 15:43
yukaCool. Do you happen to have a 16-inch model to check if it has more pstates?#asahi-dev · 2026-06-02 13:53
yukaTraces make this so much easier#asahi-dev · 2026-06-02 13:53
yukaso yeah#asahi-dev · 2026-06-02 13:53
yukabut later the port doesn't reach idle state#asahi-dev · 2026-06-02 13:52
yuka*and apply the tunables#asahi-dev · 2026-06-02 13:52
yukaAfter some more fiddling I can read the full pcie phy-ip section on t8132 amd p#asahi-dev · 2026-06-02 06:26
yukaAnd if anyone really understands how cpu_pstate_latencies.py works and why changing "off = 256" to some other number completely changes the results, please tell us :)#asahi-dev · 2026-06-02 06:25
yukalots of help from chaos_princess in getting the cpu_pstate_latencies.py to work#asahi-dev · 2026-06-02 06:24
yukahttps://github.com/yuyuyureka/linux/commits/t6031-cpufreq/#asahi-dev · 2026-06-02 06:24
yukahttps://github.com/yuyuyureka/m1n1/commits/t6031-cpufreq/#asahi-dev · 2026-06-02 06:24
yukafor anyone else wanting to experiment with cpufreq on t603[124]:#asahi-dev · 2026-06-01 10:59
yukaI published it with the changes from the weekend which make base auto-detection and chainload mostly work