#asahi-dev · 2026-10-28 06:10
evancarroll[m]I have some questions about AMX. What is the obstacle to getting that turned on? I'm a contributor to the tract project, and I see they support AMX instructions which are hand-rollled. https://github.com/sonos/tract/blob/main/linalg/arm64/apple_amx/apple_amx_mmm_f16_64x32.S.j2#asahi-dev · 2026-10-27 21:54
jannauI think the most important "fixes" are `iova_range` for dart-dcp and `stream=5` for the DCPClient#asahi-dev · 2026-10-27 21:53
jannauyuka: current diff: https://gist.github.com/jannau/674850a3d03d7e10bceefd1e13f5c5a1#asahi-dev · 2026-10-27 21:48
yukajannau: what did you fix?#asahi-dev · 2026-10-27 21:43
jannauyes, m3 / m3 pro work#asahi-dev · 2026-10-27 21:41
JamesCalligeros[m]Is this still only failing on m3 max#asahi-dev · 2026-10-27 21:31
jannauwhich would explains why nothing happens after that#asahi-dev · 2026-10-27 21:31
jannau0x__0 is type DCPEp_SetShmem, 0x_4_ is flag which was sent. 0x1__ is unknown. since we haven't seen DCPEp_SetShmem messages from DCP I guess it's an error#asahi-dev · 2026-10-27 21:28
jannauthe firmware answers the SetShmem init call with 0x140 which to my best guess is an error return#asahi-dev · 2026-10-27 21:27
yukawould be very helpful if you can check that in#asahi-dev · 2026-10-27 21:26
jannauI've fixed experiments/dcp.py enough to work to get to the same point as the linux driver.#asahi-dev · 2026-10-27 20:49
chadmedwasnt required for j314 though...#asahi-dev · 2026-10-27 20:32
svenoh, true#asahi-dev · 2026-10-27 20:27
chaos_princessprobably 14.x specific#asahi-dev · 2026-10-27 20:25
svenit's kinda odd that patch isn't required for m4 or m5#asahi-dev · 2026-10-27 20:23
jannauyes, booted to login screen#asahi-dev · 2026-10-27 20:23
yukayay#asahi-dev · 2026-10-27 20:23
jannauthat seems to do the trick#asahi-dev · 2026-10-27 20:22
yuka_recount_transition#asahi-dev · 2026-10-27 20:21
yukaAnd that one patch#asahi-dev · 2026-10-27 20:18
yukajannau: only special thing I do is disable ecores#asahi-dev · 2026-10-27 19:34
tryhard[m]jannau: maybe thats something that they got from the duo?#asahi-dev · 2026-10-27 19:34
jannauyuka: does t6034 boot macos somewhat relieable under the hypervisor? Any special boot flags or patches? t6031 panics somewhat reliably due to 2-3 timeouts#asahi-dev · 2026-10-27 18:24
sventhink different(tm)#asahi-dev · 2026-10-27 18:23
jannauwhy apple? j514c ADT has in aliases "dcp0 = /arm-io/dcp0" but the dcp node is "dcp"#asahi-dev · 2026-10-27 17:47
jannaudcp on j514s / t6030 works as well#asahi-dev · 2026-10-27 17:47
jannautoggling random smc GPIOs might break things#asahi-dev · 2026-10-27 17:29
jannaudcp on t8122 is working, not sure if I ned to add compatible or DCP FW versions checks to dcpep_cb_map_reg(). t8122 / 14.7.0 doesn't need the registers mapped but expects the mmio address as dva back#asahi-dev · 2026-10-27 17:15
chaos_princesswhat if you break into hv while in macos and do a dump of that dart, is it weird in some way?#asahi-dev · 2026-10-27 17:14
yukaI'm not sure what to truncate there since it's already only <32 bits#asahi-dev · 2026-10-27 17:13
yukaThe dvas I'm getting in the m1n1 experiment are 0x8002x000#asahi-dev · 2026-10-27 17:08
yukadart_dump.py says FOUR_LEVELS=0#asahi-dev · 2026-10-27 17:07
yukaif I read the linux code correctly?#asahi-dev · 2026-10-27 17:07
yukathat means it's locked#asahi-dev · 2026-10-27 17:07
yukaread32(u.adt["/arm-io/dart-dcp"].get_reg(0)[0]+0x208) => 0x1e7#asahi-dev · 2026-10-27 16:56
jannaum3 max#asahi-dev · 2026-10-27 16:56
jannauerr, j514m/c#asahi-dev · 2026-10-27 16:55
yukaI don't have a j514s#asahi-dev · 2026-10-27 16:55
jannauis j514s' dart locked, does it have four levels set?#asahi-dev · 2026-10-27 16:51
yukathe one on j514s does not#asahi-dev · 2026-10-27 16:51
yukathe dcp adt node on j514m has property "truncate-dvas"#asahi-dev · 2026-10-27 16:15
svenlooks like that was caused by a fiq storm because the HACR trap for PMCR0 no longer works#asahi-dev · 2026-10-27 16:15
svenyuka: https://github.com/AsahiLinux/m1n1/pull/742/changes/98a0e98d139da43f282a1b6a12c43ccbf8e00d7b has fixed the hangs after I login fwiw#asahi-dev · 2026-10-27 15:16
svenor only increase the heap to something bigger for non-macho payloads in run_guest.py#asahi-dev · 2026-10-27 15:07
jannauI bumped the heap to 428MB. I guess I'll add a command line parameter so that run_guest_kernel.sh can request a size for my huge boot images#asahi-dev · 2026-10-27 15:05
svenjust patching that check out also works#asahi-dev · 2026-10-27 15:05
sventhat stupid thing is the reason I had to move the shadow pagetables around twice :D#asahi-dev · 2026-10-27 15:04
jannauyep, that was it#asahi-dev · 2026-10-27 15:04
sven*bit more#asahi-dev · 2026-10-27 15:04
sveniirc we have a bit than 100 MiB of headroom until it hits that check