#asahi-dev · 2026-02-07 09:32
nickchanjannau: though what do you mean by trap control #asahi-dev · 2026-02-07 09:30
noopwafelI have an #ifdef GOING_TO_BOOT_LINUX wrapped around those in m1n1 atm..#asahi-dev · 2026-02-07 09:27
nickchanwith VHE schematics program AGTCNTRDIR_EL1 and AGTCNTRDIR_EL12 to identical values in m1n1 EL2 and it will be consistent for kvm#asahi-dev · 2026-02-07 09:24
jannaunickchan: I meant we need to ensure the EL1 and EL2 state is consistent for KVM#asahi-dev · 2026-02-07 09:23
nickchanand if AGTCNTRDIR_EL1 and AGTCNTRDIR_EL12 (from EL2 pov) is consistent then the state of HCR_EL2.TGE should not matter for counter#asahi-dev · 2026-02-07 09:22
nickchanguest can't use the register then#asahi-dev · 2026-02-07 09:21
svensomeone with a m4 should check if we can write to those or if they already have the correct value#asahi-dev · 2026-02-07 09:21
nickchanjannau: i think kvm can ignore those registers since it sets TIDCP#asahi-dev · 2026-02-07 09:20
jannausigh, yes. KVM needs AGTCNTRDIR_EL12. I was just thinking of m1n1 as hypervisor and there _EL1 is enough since the hypervisor loads m1n1 anyway#asahi-dev · 2026-02-07 09:17
nickchanprogram both AGTCNTRDIR_EL1 and AGTCNTRDIR_EL12 and it should be fine though#asahi-dev · 2026-02-07 09:15
nickchanFor linux the EL0 and EL1/2 counter frequency should be aligned#asahi-dev · 2026-02-07 09:04
jannauI finally found out what BIT(0) of AGTCNTRDIR does. It is an additional EL0 trap control for CNT accesses #asahi-dev · 2026-02-07 04:51
integralpilotooh looks like fuchsia is finally sending some stuff over UART#asahi-dev · 2026-02-06 21:56
jannauI was confused by the commit date in the message and thought this was an old report. I checked that fixup was only added after that#asahi-dev · 2026-02-06 21:55
sven*thought#asahi-dev · 2026-02-06 21:54
svensaw it earlier today and was confused for a moment as well because i though i had already fixed it :/#asahi-dev · 2026-02-06 21:54
jannausven: argh, I can't read. good catch#asahi-dev · 2026-02-06 21:44
svenlooks like the usb curse got to you now@#asahi-dev · 2026-02-06 21:40
noopwafelhm, wrote new BIST dance and .. it also fails waiting for a status register :)#asahi-dev · 2026-02-06 20:33
Meleodyboth of which I'll have to compile from source#asahi-dev · 2026-02-06 20:26
Meleodylooks like the choice is between pytorch with opencl or vulkan#asahi-dev · 2026-02-06 20:21
jannauand OpenCL. the GPU has a mayor quirk though. There is no preemption for compute jobs#asahi-dev · 2026-02-06 20:21
chaos_princessopencl is there, may be more appropriate#asahi-dev · 2026-02-06 20:20
jannauthe ane driver programs DART directly instead of using linux' DMA API#asahi-dev · 2026-02-06 20:20
MeleodyVulkan and OpenGL support rn, correct?#asahi-dev · 2026-02-06 20:18
chaos_princessjust use the gpu? #asahi-dev · 2026-02-06 20:17
MeleodyFine, no Neural Engine for me ig :(#asahi-dev · 2026-02-06 20:16
jannauand open user user space is not just the program which loads pre-compiled microcode but the compiler#asahi-dev · 2026-02-06 20:15
jannauhttps://docs.kernel.org/accel/introduction.html references https://docs.kernel.org/gpu/drm-uapi.html#open-source-userspace-requirements with "... All of that is the same for the accel subsystem"#asahi-dev · 2026-02-06 20:13
chaos_princessthere is most likely no need to consider sme on m4, as we will get that for free from generic arm64#asahi-dev · 2026-02-06 20:12
chaos_princesssounds pretty gpu-like :P#asahi-dev · 2026-02-06 20:12
chaos_princessit is a thingy that does dma and computation based on a program you load into it#asahi-dev · 2026-02-06 20:12
MeleodyAnother thing to consider is SME on the M4. But I guess it's gonna be a long road until then.#asahi-dev · 2026-02-06 20:12
chaos_princessbecause it is in drm/accel#asahi-dev · 2026-02-06 20:11
MeleodyI can work without an ecosystem with the tools that are there. Also the Neural Engine isn't a gpu so why should it follow gpu rules just cause it does matrix stuff?#asahi-dev · 2026-02-06 20:08
svenand if this follow the gpu rules you need an open-source userspace with a stable UABI before the kernel module will be accepted#asahi-dev · 2026-02-06 20:08
svenyeah, and that#asahi-dev · 2026-02-06 20:07
chaos_princessyour main issue is that there is no ecosystem to use those things on linux#asahi-dev · 2026-02-06 20:07
svenfrom what i remember the current driver has quite a bit of hacks, especially regarding the iommu handling#asahi-dev · 2026-02-06 20:07
Meleodytbf the speedup depends on the application. I'll watch the linux kernel space more closely and see how it could be upstreamed.#asahi-dev · 2026-02-06 20:05
chaos_princessas ane is only capable of running the encoder part of the whisper model, while the runtime is dominated by the decoder#asahi-dev · 2026-02-06 20:04
chaos_princessjfyi, with my experimentation, on a 1-hr audio file, using ane saves like a second or two compared to gpu only#asahi-dev · 2026-02-06 20:04
svenit's simple really: if you want to get this into the tree work on getting it upstream. it'll end up in our tree soon after it hits -next#asahi-dev · 2026-02-06 20:03
Meleodypractical stuff? https://github.com/eiln/ane-ex & https://github.com/ggml-org/whisper.cpp/pull/1021. Technically? A way to run hardware microcode for ML Models generated on Mac OS#asahi-dev · 2026-02-06 20:00
chaos_princesswhat does it provide atm?#asahi-dev · 2026-02-06 19:58
MeleodyAnd I don't mean to say that it should be unmaintained, it still needs work for M2 and stuff.#asahi-dev · 2026-02-06 19:57
MeleodyI just think being pragmatic and sticking to the scope of what the ANE Driver provides atm would be benficial imo.#asahi-dev · 2026-02-06 19:51
chaos_princessuserspace = "Implementing a whole ML Framework"#asahi-dev · 2026-02-06 19:46
MeleodyOne last question regarding the AME: what counts as userspace? Implementing a whole ML Framework is a lot of work for very little gain, but never upstreaming the simple parts allowing basic access is just a shame imo.