#
#asahi-dev · 2026-10-28 08:15
yukalet's see if dcpext also works
#asahi-dev · 2026-10-28 08:15
yukaAnd I also changed the dependencies of dart / dcp to _fe since that's now the "top-level" one
#asahi-dev · 2026-10-28 08:14
jannauack. that's the minimal change that makes it work here as well
#asahi-dev · 2026-10-28 08:13
yukajust moved the min-state from _fe to _cpu
#asahi-dev · 2026-10-28 08:13
chadmedlets go
#asahi-dev · 2026-10-28 08:13
yukagot it working as well
#asahi-dev · 2026-10-28 08:13
yukaNice!
#asahi-dev · 2026-10-28 08:12
jannauno, we turning something off too much. pd_ignore_unused is not required
#asahi-dev · 2026-10-28 08:11
chadmedso are we just not turning something on properly
#asahi-dev · 2026-10-28 08:10
jannauon t6031
#asahi-dev · 2026-10-28 08:10
yukaon t603[124]?
#asahi-dev · 2026-10-28 08:10
jannauyuka: dcp works in linux with pd_ignore_unused and "apple,min-state = <4>;" sprinkled over more disp power-states
#asahi-dev · 2026-10-28 08:09
yukaHope the setshmem flag difference helps if you want to try again
#asahi-dev · 2026-10-28 08:06
jannauI might have been misguided by crashlog buffer though
#asahi-dev · 2026-10-28 08:06
jannauthe iova allocator from dart.py. the default overlaps with already allocated iova
#asahi-dev · 2026-10-28 08:04
yuka(for the experiment)
#asahi-dev · 2026-10-28 08:04
yukaI also needed iova_range=(0x007800000, 0x40000000), where did that come from?
#asahi-dev · 2026-10-28 07:53
yukaSo I can compare experiment and linux
#asahi-dev · 2026-10-28 07:53
yukastill, I now have a working-ish experiment?
#asahi-dev · 2026-10-28 07:52
sven:(
#asahi-dev · 2026-10-28 07:51
yukathat's why I didn't see it before, there was no difference!
#asahi-dev · 2026-10-28 07:50
yukayes: linux: SetShmem [...] FLAG=0x4
#asahi-dev · 2026-10-28 07:49
yukawill trace it again
#asahi-dev · 2026-10-28 07:49
yukaSo then the kernel is failing in another way
#asahi-dev · 2026-10-28 07:48
yukathe kernel already sets flag to 4
#asahi-dev · 2026-10-28 07:48
yukahuh
#asahi-dev · 2026-10-28 07:48
yukaOh
#asahi-dev · 2026-10-28 07:46
yukanow need to find the place where this message is actually sent in the kernel
#asahi-dev · 2026-10-28 07:45
JamesCalligeros[m]Not
#asahi-dev · 2026-10-28 07:45
JamesCalligeros[m]Why cant they just
#asahi-dev · 2026-10-28 07:45
JamesCalligeros[m]Eyyyy
#asahi-dev · 2026-10-28 07:43
svenjust a little bit ;)
#asahi-dev · 2026-10-28 07:43
yukathis looks better right? ;)
#asahi-dev · 2026-10-28 07:43
yukaexperiment looks like this now: https://termbin.com/823j
#asahi-dev · 2026-10-28 07:42
chadmedauuughhhhh
#asahi-dev · 2026-10-28 07:42
yukanow trying in linux
#asahi-dev · 2026-10-28 07:42
yukaexperiment: SetShmem [...] FLAG=0x0
#asahi-dev · 2026-10-28 07:41
sven🍿
#asahi-dev · 2026-10-28 07:41
yukamacOS: SetShmem [...] FLAG=0x4
#asahi-dev · 2026-10-28 07:41
chadmeduh oh
#asahi-dev · 2026-10-28 07:41
yukaI don't know how I missed this before
#asahi-dev · 2026-10-28 07:41
yukaoeeeehhh
#asahi-dev · 2026-10-28 07:29
yukajannau: nice, now the logging also works in the experiment
#asahi-dev · 2026-10-28 06:38
JamesCalligeros[m]Its a proprietary isa with an architecture manual and standard extensions that are expected to be used. Apple are free to do whatever they want in their walled garden per their architecture licence, we however are not.
#asahi-dev · 2026-10-28 06:16
svenpeter[m]upstream has already rejected the TSO bit, AMX is much more invasive
#asahi-dev · 2026-10-28 06:15
evancarroll[m]Arm itself is a propeietary ISA... Shouldn't the only quesiton be whether or not it's documented and in use? https://github.com/corsix/amx
#asahi-dev · 2026-10-28 06:12
JamesCalligeros[m]We will never enable AMX. It is a proprietary ISA extension that has no business in upstream Linux. Newer machines support SVE and SME, which are the replacements for AMX.
#asahi-dev · 2026-10-28 06:10
evancarroll[m]* AMX instructions on OSX which are
#asahi-dev · 2026-10-28 06:10
evancarroll[m]I mean to say, we don't actually need the kernel to support AMX's ISA. We just need it turned on and the hand rolled instructions in tract would just work?
#asahi-dev · 2026-10-28 06:10
evancarroll[m]If Asahi set the register to enable AMX (even if only with a kernel boot param), and we submitted a patch for lazy context-switching (preserving the coprocessor's state around context switching only when the amx instruction where hit), what else is required?

plus ›