Repository navigation
accel/ane: T8142 (M5) Neural Engine, enabled on J815 - #1883
Merged
DjDeveloperr merged 8 commits intoNov 14, 2026
Merged
Conversation
The M5's ANE is the H17 engine (ADT "ane,t8132exclave", ane-type 0x210) with four DAPF records, T8140's interrupts and a preload in reserved memory. Its PWGATE is a window of its own, so it takes the power domains the T6050 form does. Assisted-by: Claude Opus 5.5 Signed-off-by: Nick <nekreshen@yandex.ru>
The M5's ANE DARTs are T8140's cold ANE DARTs at revision 0x204. Assisted-by: Claude Opus 5.5 Signed-off-by: Nick <nekreshen@yandex.ru>
The M5's ANE is the H17 engine again (ADT "ane,t8132exclave", ane-type 0x210), exclave-assigned and preloaded by iBoot. From the T8142 ADT: ANE_SYS at PMGR +0x2e8 (pmgr device 73), ANE_CPU +0xc000 and ANE_TDBASE +0xc008 as on T8140, four DAPF records, and a preload of TEXT 0xbc000 and DATA 0x2b0000 at 0x100_0000_0000 / 0x100_000b_c000. Its PWGATE window, scratch cells, DART revision and TEXT digest are not measured yet. Give ane_t8142_platform a zero TEXT pin that never matches: with a loader hand-off the driver only logs it and leaves the engine alone. The hand-off check now hashes the handed-over TEXT even when its layout is not the platform's, and logs that layout next to the digest, so one boot of a new SoC or firmware reports both. Assisted-by: Claude Opus 5.5 Signed-off-by: Nick <nekreshen@yandex.ru>
Under aura on a J815 with the macOS 27.0 (26A428) stub, iBoot preloads the ANE TEXT 0xc0000 and DATA 0x2ac000 at 0x100_0000_0000 and 0x100_000c_0000, T8140's layout (the 27.2 host's ADT shows 0xbc000 and 0x2b0000). Use it, and keep the measured TEXT digest next to the zero pin until the PWGATE window, scratch cells and DART revision are known. Assisted-by: Claude Opus 5.5 Signed-off-by: Nick <nekreshen@yandex.ru>
The M5's three ANE DARTs are T8140's: 16 streams, the exclave's streams inherited from iBoot, SID0 the only one the kernel attaches. They are revision 0x204 (PARAMS_8 0x2a2a0204, measured on a J815) where T8140's are 0x203. Give apple,t8142-ane-dart the cold ANE profile and check for its own revision. Assisted-by: Claude Opus 5.5 Signed-off-by: Nick <nekreshen@yandex.ru>
Measured under aura's hypervisor on a J815 running macOS 27.2.1 (26B131), with a cold model load, the power-up, the first program and the power-down traced: - PWGATE is PMGR +0x27640 (the ADT's ane reg[2] + 0x3640): request +0xc (0 release, 3 gate), acknowledge +0x30, as on T8140. - The scratch cells start at +0x1880028, as on T6050. - The DARTs are revision 0x204 (PARAMS_8 0x2a2a0204), 16 streams; the ADT gives SIDs 1-7 to the exclave and bypasses 10 and 13. - Four DAPF records, the DT's. - macOS powers the engine down in T8140's order: sub-block, PWGATE request 3, acknowledge 0, CPU, then ANE_SYS. So T8142 runs as T8140 does: EXCLAVE_MODE_STOP before a clean stop, the engine parked until its first request and cycled cold across system sleep, the preload pinned by the 26A428 TEXT digest measured on a J815. Tested on a J815 (7.3.0, integration/aurora, iBoot preload of the 26A428 stub): the engine binds and parks behind its three cold DARTs, the firmware boots on the first request, and an H17 program from Apple's compiler loads and its process is created and torn down (H17 LOAD, CREATE, TERMINATE and UNLOAD all status 0, with the H17 HWX validator change for zin_ane_compiler 10.205 programs). The ANE KUnit suites pass but for program_add128_slots_test, which fails on integration/aurora already. A frame through the engine is not yet compared against a reference. Assisted-by: Claude Opus 5.5 Signed-off-by: Nick <nekreshen@yandex.ru>
Describe ANE_SYS (pmgr device 73, +0x2e8), the three ANE DARTs (0x421800000, 0x421820000, 0x421840000, IRQ 752, vm range 0x0-0x400_0000_0000) and the engine: its registers, the PMGR window and PWGATE (PMGR +0x27640), IRQs 751 and 764, the firmware IOVAs and four DAPF ranges, disabled. Assisted-by: Claude Opus 5.5 Signed-off-by: Nick <nekreshen@yandex.ru>
Assisted-by: Claude Opus 5.5 Signed-off-by: Nick <nekreshen@yandex.ru>
m0d16l14n1
marked this pull request as ready for review
November 14, 2026 06:47
DjDeveloperr
approved these changes
Nov 14, 2026
DjDeveloperr
left a comment
Contributor
There was a problem hiding this comment.
Reviewed. The T8142 platform data, match entries and DT are new and J815-only. The two shared hunks:
apple-dart.c:cold_aneand the 0x204 cold revision apply only toapple,t8142-ane-dart. T8140 keeps 0x203, and T6050 DARTs never take the cold path.ane_fw_cold.c: the hand-off now hashes TEXT even when the preload layout differs.text_okis stillranges_valid && digest == pin, so acceptance is unchanged on T8140 and T6050; a mismatch only adds one log line.
The reference-checked execution and the s2idle cycle are good evidence. The contiguous-scratch CMA limit for the 960x540 program is a fair follow-up.
dt-checks and host-tests are green on this head. Base drift doesn't touch these files. Approving and merging. Heads-up: #1880 and #1881 append at the same places in t8142-j815.dts and t8142-peripherals.dtsi, so they will need a rebase after this.
DjDeveloperr
merged commit Nov 14, 2026
2bc6694
into
aurora-silicon:integration/aurora
2 checks passed
This was referenced Nov 14, 2026
This was referenced Nov 14, 2026
DjDeveloperr
pushed a commit
to aurora-silicon/aura
that referenced
this pull request
Nov 17, 2026
/arm-io/ane on the M5 MacBook Airs (J813, J815) is pre-loaded by iBoot with a __TEXT and a __DATA segment, as on T8140. List apple,t8142-ane so an M5 ANE node gets both reserved and its firmware IOVAs. Tested on J815 (MacBook Air 15" M5) with the macOS 27.0 (26A428) and 27.2 (26B123) stubs, on v1.0.0b13.1 plus this change: the kernel's T8142 ANE driver (aurora-silicon/linux#1883) finds both regions at 0x100_0000_0000, verifies the TEXT and runs an H17 program whose output matches macOS's (1080p frame, luma 100% within 1). Signed-off-by: Nick <nekreshen@yandex.ru> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The M5 (T8142) Neural Engine on the H17 staged backend, as T8140 runs it, and enabled on J815.
apple,t8142-ane(ane-type 0x210, four DAPF records, T8140's interrupts, a preload in reserved memory, the T6050 form's power domains) andapple,t8142-ane-dart.apple,t8142-ane-dart, at their own revision 0x204 (T8140's are 0x203).Tested on J815 (7.3.0, integration/aurora, iBoot preload of the 26A428 stub): the engine binds and parks behind its three cold DARTs, the firmware boots on the first request, and an H17 program from Apple's compiler loads; its process is created and torn down (
H17 LOAD,CREATE,TERMINATE and UNLOADall status 0, with #1882). Execution checked against a reference: the 960x540 program of an open camera-denoise network (compiled for H17 by macOS's ANE compiler) run over a seeded 1920x1080 NV12 frame gives luma within 1 of the macOS CPU reference output on 100.00% of pixels, with chroma untouched (the raw input matches the reference on only 14.5%, so a run that does nothing cannot pass); about 26-32 ms per frame. Across system sleep: with the engine in use, one s2idle cycle (about 23 s): the firmware is stopped and the engine powered down for the sleep, parked on resume, brought up cold on the next request, and the same frame again matches the reference. The 640x360 program also runs in the live camera pipeline. The ANE KUnit suites pass exceptprogram_add128_slots_test, which already fails on integration/aurora (see #1790). checkpatch --strict clean; CHECK_DTBS clean for J815, J813 and J704 on the touched schemas.Open: a 960x540 program's 63 MiB scratch is one physically contiguous allocation (
ane_fw_alloc_mem), which the default 64 MiB CMA could not give (ENOMEM); a larger CMA works around it for now, mapping scattered pages behind the DART would fix it. Tested withcma=256M.Part of the J815 (MacBook Air 15-inch, M5, T8142) bring-up: board results are summarised in discussion #135.
Assisted-by: Claude Opus 5.5