#asahi-dev · 2022-09-13 16:15
arndright, good idea#asahi-dev · 2022-09-13 15:46
arndI would also skip the 'default y' and just put that in the defconfig files, the only reason for default-enabling those options is that it would otherwise break existing defconfigs that don't have it enabled#asahi-dev · 2022-09-13 15:43
arnddrivers/platform/surface/ uses 'depends on ARM64 || X86 || COMPILE_TEST', which seems like a good idea in general, even if it doesn't make much difference#asahi-dev · 2022-08-16 08:54
arndpjakobsson: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=93929fb6c4ef6211e96679f060c842f35fac7710#asahi-dev · 2022-08-16 08:45
arndbut I have no experience actually doing that#asahi-dev · 2022-08-16 08:45
arndpjakobsson: I think the idea to use tracepoints on each mmio access is generally useful, if you can filter the results in a useful way to get the trace you want. What I don't know is whether the implementation we have is good enough for that. My hope would be that one can use bpf (https://lwn.net/Articles/683504/) for filtering the results to get just the right ones#asahi-dev · 2022-08-16 08:39
arndif you have any ideas for improvements, I'm sure others can benefit from that as well#asahi-dev · 2022-08-16 08:37
arndpjakobsson: so does the code we merged work for this usecase?#asahi-dev · 2022-08-16 08:25
arndUnfortunately the ideas that I had for making it more general tended to also break their use case.#asahi-dev · 2022-08-16 08:24
arndmarcan: I'm also not sure if the CONFIG_TRACE_MMIO_ACCESS implementation is generally that useful. This has been going for a while, and as I understand, the authors mainly need it as a debugging tool for qcom chips that completely hang when a broken driver accesses an MMIO location that is owned by EL3. We'll see if someone can make it work for anything else#asahi-dev · 2022-05-13 16:04
arndI still want to eventually turn on all those options in mainline, so I can just build a defconfig kernel and not think about it much, but there is not much point until enough drivers are there#asahi-dev · 2022-05-13 16:03
arndmaz: makes sense. I wasted a ton of time just finding the right kernel config. I had started with arm64 defconfig and then turned on stuff one option at a time, which meant a lot of reboots until I had something working ;-)#asahi-dev · 2022-05-13 15:52
arndmaz: I meant just to get you started with setting up user space. Once it works in the VM, you can obviously boot the same kernel natively from proxyclient, and then move to u-boot the last step. I found the vm proxyclient to be super helpful just getting started when not following the normal install#asahi-dev · 2022-05-13 15:31
arndmaz: have you tried just booting a kernel in a m1n1 vm? that's how I'm still booting at the moment, skipping the u-boot step#asahi-dev · 2022-05-07 19:08
arndI wonder why we never did that for armada8k#asahi-dev · 2022-05-07 19:07
arndjannau: oh, that's great. It's nice when the .dts syntax actually works in order to get something done rather than making it harder ;-)#asahi-dev · 2022-05-06 21:04
arndpushed out with the latest soc changes now#asahi-dev · 2022-05-06 19:09
arndsven: can you have a quick look at this fixup? https://www.irccloud.com/pastebin/EtJPzidT/#asahi-dev · 2022-05-02 07:49
arndsure#asahi-dev · 2022-05-02 07:45
arndsven: if the changes are very small, just reply that you've folded in the obvious fix, and then send the pull request when you feel it's ready#asahi-dev · 2022-04-28 10:05
arndmarcan: do you have any other drivers/soc/ contents for 5.19? If not, I guess sven can send the subsect pull request to soc@kernel.org directly with your ack#asahi-dev · 2022-04-28 10:02
arndit might be nice to have the rtkit+sart portion of the nvme branch in the soc tree anyway, even if nvme is the only thing that needs it#asahi-dev · 2022-04-11 12:49
arndin the gcc source tree, run ./contrib/download_prerequisites to make it build a local copy of the mpc/mpfr/isl/gmp libraries instead of the distro version#asahi-dev · 2022-04-11 12:46
arndI don't know who jmesmon is, but it looks like nathanchance has a newer version at https://github.com/nathanchance/buildall#asahi-dev · 2022-04-11 12:44
arndthe toolchains are based on segher's 'buildall' scripts, I need to see if I can find a copy#asahi-dev · 2022-04-11 12:43
arndit might be enough to copy libc.so.6 from any other distro, but I haven't tried#asahi-dev · 2022-04-11 12:42
arndhttps://mirrors.edge.kernel.org/pub/tools/crosstool/files/bin/arm64/ has toolchains that should work on any distro, as long as you have glibc, using those should make it easier to compare across systems#asahi-dev · 2022-04-11 12:40
arndj`ey: clang is usually some 10% to 20% slower than gcc for a defconfig build, but it's more sensitive to options that lead to larger indirect header inclusions#asahi-dev · 2022-04-11 12:36
arndthe time it takes for building a kernel can differ hugely based on a single Kconfig option, such as CONFIG_DEBUG_INFO, the exact toolchain version, or how the compiler was built#asahi-dev · 2022-04-11 11:46
arndjannau: on this many cores, the total runtime is not all that meaningful because half the build time is spent in single-threaded work like linking or parsing the Makefiles, it's often better to compare CPU time (user+system from /bin/time, or output from perf stat) to see how much work the CPUs got done#asahi-dev · 2022-04-11 10:35
arndinterestingly it goes down to 886 GFLOPS if I use all 20 cores including the small ones, but I think that's just blis/gemm not being aware of big/little cores, and making the big ones wait#asahi-dev · 2022-04-11 10:34
arndit scales almost perfectly now: 591 GFLOPS for either set of 8 firestorm cores, 1171 GFLOPS for all 16#asahi-dev · 2022-04-11 10:25
arndjannau: what is the maximum frequency under constant load? If the m1 max gets throttled to the same 2GHz after a while, that still works out#asahi-dev · 2022-04-11 10:21
arndwith the wrong number, CPU 12 dropped down to 18 GFLOPS...#asahi-dev · 2022-04-11 10:20
arndI got a typo in cpu_p00_d1, that needs to be 4, not 3#asahi-dev · 2022-04-11 10:18
arndnot sure if marcan ends up redoing it all anyway#asahi-dev · 2022-04-11 10:18
arndjannau: should I send you a proper patch, or do you just want to fold it in?#asahi-dev · 2022-04-11 10:17
arndhttps://www.irccloud.com/pastebin/DCvvTxzq/#asahi-dev · 2022-04-11 10:17
arndI had to change the reg-names as well, but now it seems to work#asahi-dev · 2022-04-11 10:16
arndjannau, right, I just tried that locally#asahi-dev · 2022-04-11 10:02
arndso it thinks that the cores in the second cluster always run at the same clk as the corresponding ones in the first one. looking at cur_freq confirms that#asahi-dev · 2022-04-11 10:00
arnd2 3 4 5 12 13 14 15#asahi-dev · 2022-04-11 10:00
arnd$ cat /sys/devices/system/cpu/cpu12/cpufreq/affected_cpus#asahi-dev · 2022-04-11 09:50
arndI guess cpuinfo_cur_freq shows what the kernel has requested, not what the hardware actually does#asahi-dev · 2022-04-11 09:49
arndok, that explains it then#asahi-dev · 2022-04-11 09:49
arndj`ey: what was the issue? I see the firestorm cores running at 3228000 HZ in /sys/devices/system/cpu/cpu12/cpufreq/cpuinfo_cur_freq during the test#asahi-dev · 2022-04-11 09:42
arndhttps://www.irccloud.com/pastebin/S2KGO55M/#asahi-dev · 2022-04-11 09:42
arndGFLOPS on the other eight#asahi-dev · 2022-04-11 09:42
arndI ran into some funny performance issue on the second half of the T6002. I noticed that a BLAS/GEMM benchmark is a lot slower on 10 cores of the M1 Ultra than an M1 Max MBP. With some more investigation I found that the first 10 cores are about the same, but the second set of firestorm cores appears slower. Doing the test individually on each core shows 30 GFLOPS on any icestorm, 90 GFLOPS on the first eight firestorm cores and 60