#
✕ nick:alexaltea
#asahi-re · 2025-10-04 18:27
alexalteachaos_princess, ok that makes sense, thanks. so far drivers expose /dev/accel/ane0..3, I'll have libane pick aneN via round robin, unless explicitly stated
#asahi-re · 2025-10-04 16:12
alexalteaalso, I'm assuming that each .hwx model only needs ONE ane node to run... I haven't seen anything resembling sync primitives so hopefully there's no cooperation between nodes
#asahi-re · 2025-10-04 16:09
alexalteaand if so, any insights on how such engines are selected in macos to run a task? are engines selected by macos system libs? or do have apps a say in which ones to use?
#asahi-re · 2025-10-04 16:08
alexalteahas anybody encountered this pattern in other devices? (multiple independent engines)
#asahi-re · 2025-10-04 16:07
alexalteaeiln documented this at https://github.com/eiln/ane/blob/33a61249d773f8f50c02ab0b9fe5b2db58346751/docs/ane-boards.md
#asahi-re · 2025-10-04 16:07
alexalteaonly one question remains: some apple machines have several ane nodes (ane0, ane1, ane2, ane3), which seem completely independent from each other (I guess each have their own DARTs too?)
#asahi-re · 2025-10-04 16:07
alexalteaI've added m2 (h14) support to eiln's ane drivers, which were m1 (h13) only. I've made some changes to the uapi and userland library, so that it can load models directly from .hwx files without prior conversion to the custom .anec format