Skip to content

accel/ane: bring up J616c 25F84 inference diagnostics - #805

Closed
astaff wants to merge 63 commits into
aurora-silicon:integration/aurorafrom
astaff:t6041/j616c-ane-25f84
Closed

astaff wants to merge 63 commits into
aurora-silicon:integration/aurorafrom
astaff:t6041/j616c-ane-25f84

Conversation

@astaff

@astaff astaff commented Nov 1, 2026 •

Copy link
Copy Markdown
Contributor

Why

The M4 Max's Neural Engine is the same H16 generation as the M4 Pro's, whose cold-boot stack is now upstream (#720). J616c needs its device-tree description and support for the resident 25F84 firmware image, and the bring-up produced a set of bounded, default-off diagnostics for startup ordering, report retirement, interrupt and polling cadence and recovery, with KUnit and selftests.

What (current branch)

  • t6041-j616c-ane.dtsi and the SoC file: the J616c ANE, its DARTs and power states.
  • accel/ane: 25F84 resident-image support and the diagnostics listed above; pmdomain/apple: PMP report, DRAM and ANE performance-level requests with tests.

Status

Draft, not yet mergeable. The branch imports the H16 series as it was before #720 merged; upstream has since reworked the driver substantially (new program, lifetime and power-management files), so the diagnostics need porting onto the merged driver as a smaller series, and the device-tree part needs re-checking against the fuller ANE DART description now in t604x.dtsi. The device-tree commits will be re-stacked on #525 first; the driver diagnostics follow once they are rebuilt on the current H16 code.

astaff and others added 30 commits October 28, 2026 17:11
Reuse the existing SoC and PMGR descriptions as the shared T604x base.
The J616s compiled tree keeps every node and property unchanged.

Assisted-by: LLM
Describe the Max CPU, OPP and register-window differences on the shared
base. Preserve the qualified J616c node and property values.

Assisted-by: LLM
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Move identical register-window constructors into a family helper.
The T6040 and T6041 initialization-data checks retain their exact values.

Assisted-by: LLM
Select a static 25F84 profile and admit the normalized loaded image
measured in three qualification boots. Remove exploratory overrides
and unknown-image admission from the support path.

Assisted-by: LLM
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Keep the reason-0 and cancellation handling scoped to 25F84, with
serial rendering explicitly opt-in. Concurrent-render firmware restarts
and complete recovery remain unresolved.

Assisted-by: LLM
Add gated 26.5.2 firmware acceptance and headerless powerlog parsing
for the T6041 display controller.

Assisted-by: LLM
Wait for the flip notification before completing commit cleanup so
framebuffer ownership follows scanout completion.

Assisted-by: LLM
Give bootarg access width its own hardware setting and patch through
one aligned buffer. Preserve surrounding bytes and verify changed words.
Synthetic MMIO checks cover malformed tables, bounds and failed readback.

Assisted-by: LLM
Describe the six amplifiers and headphone jack in the J616c board
DT. Stop playback when the protection lease is lost.

Assisted-by: LLM
Limit admission to the qualified J616c machine and firmware layout.
Apply the capability per device without allocating or changing the
shared variant descriptor.

Assisted-by: LLM
Register the watchdog against its platform device and apply the
configured nowayout policy.

Assisted-by: LLM
Build the production initialization objects with their soft-float types
and compare constants with explicit capture inputs. Keep T6040 per-chip
calibration validation separate; include no captures or local defaults.

Assisted-by: LLM
m1n1 compares DT CPUs with its physical MPIDR enumeration. Appending the Max cores to the Pro topology put CPU 9 out of order and stopped boot before kernel entry.

Define the full family topology in hardware order and remove the extra cores from J616s. Check the compiled DTB order for both SoCs.

Assisted-by: LLM
Keep all shared firmware profiles at TA 0x1c890 = 0 and enable mode 1
only in T6041's admitted 25F84 build. Pass that device/build policy
through RenderArgs without changing register ordering, merge-buffer
writes, fragment programs or serial-render admission.

Add KUnit coverage for the policy boundary and constructor output,
including SCS array shape and unchanged fragment/G17P arrays.

Motivated by AxxzyWasTaken's separate J616C/25G83 implementation results:
aurora-silicon#525 (comment)
Those results do not validate this branch or its 25F84 firmware; the
cluster merge-buffer explanation remains a hypothesis.

Validation: patch application/reversal, whitespace and source-scope checks.
Not yet compiled, KUnit-executed or hardware-tested. Keep the existing
asahi.t6041_serial_render=1 workaround pending the concurrent-render test
matrix documented in the PR.
drivers/accel/ane, its uapi header, documentation, binding and selftests
at e40f774, and the additive pmp-report ANE/DRAM providers they link
against. Nothing else from aurora-wip.

Assisted-by: LLM
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… ane)

Keep the T604x ANE node, its three DARTs and the ANE PMGR domains on T6041
and describe the J616c ANE with the J616s windows, DAPF records, firmware
IOVAs and five power domains, checked against this board's ADT. Two
disabled firmware reservations wait for the boot loader, which fills them
from this boot's ane0 segment-ranges and enables the chain. ANE_SYS's
fabric parents are critical, on from iBoot and not in the T6041 tree.

The DAPF bounds are one 64-bit cell group: /bits/ 64 covers only the first
<...> of a list.

Assisted-by: LLM
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…diagnostic, workspace ane)

With h16_identity=observe and h16_stage=preflight only, a preload whose
identity is refused is logged, its TEXT and DATA snapshots are published in
debugfs ane_h16/{text,data} until module exit, and preflight runs anyway.
Preflight reads power, CPU and DART/DAPF state and writes nothing to the
engine. Other stages refuse observe.

Assisted-by: LLM
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AGCaKaz9NWJpjHQ9RwzeMj
…ace ane)

A table of preload layouts: 26A428 as before, and 25F84 (macOS 26.5.2),
whose patchbay framing and tunables interval sit 0x3f80 further into DATA.
ane_h16_preload_identify returns the matching build; the backend accepts
either and keeps the 26A428 boot values until a 25F84 stage disagrees.
The 25F84 digest is unknown until a live image is observed, so it matches
nothing yet.

Assisted-by: LLM
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AGCaKaz9NWJpjHQ9RwzeMj
rg-ane-00-preflight observed the live J616c image (macOS 26.5.2 25F84):
TEXT and DATA match the firmware file except the mutable patchbay cells
(_rtk_patchbay, DATA+0x9998), _rtk_tunables (DATA+0x1bc00, 0x6a0 bytes)
and the iBoot-filled TEXT ce_env; record its identity digest.

A 25F84 build profile carries the 26A428 boot values, which preflight
confirmed (power, CPU control 0, RVBAR 0x10000000001), plus this board's
DAPF facts: the DART word at 0x8 reads 0x2a2a0202 and dart-ane0 has one
DAPF instance (ADT reg 3, DAPF LLT), so only it carries records. The
backend selects the build the identity matched.

Assisted-by: LLM
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AGCaKaz9NWJpjHQ9RwzeMj
…rkspace ane)

The H16 backend allocates its private firmware surfaces cacheable on both
sides and refuses a noncoherent client before releasing the firmware CPU.
The M1/M2 ANE runs with cacheable DART descriptors and cached CPU mappings
(accel/ane map_mode 3, buffer-reuse checks). rg-ane-03-boot tests it here.

Assisted-by: LLM
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AGCaKaz9NWJpjHQ9RwzeMj
…nostic, workspace ane)

With h16_job=1 and a completed startup stage, one write to debugfs
ane_h16/job runs records in order: MAP a zero-filled private surface at a
DVA and copy data in, WRITE or FILL a mapping, CMD (publish a host-built
command on a channel, wait for its slot, copy the reply back; program and
process handles from earlier replies can be filled in), EVENT (consume
IO_T2H reports up to the terminal event of a UUID) and PEEK. ane_h16/result
returns one record per step with status and elapsed nanoseconds. The host
builds every firmware command; this only maps, sends and waits.

Assisted-by: LLM
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AGCaKaz9NWJpjHQ9RwzeMj
…rkspace ane)

The transport wait relaxes 1 ms per poll, which made every job command take
at least ~1 ms and hid execution time. Job commands and event waits now poll
every 10-20 us while servicing IRQ status, shared allocations, the mailbox
and DART faults as ane_fw_transport_wait does.

Assisted-by: LLM
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AGCaKaz9NWJpjHQ9RwzeMj
…e ane)

rg-ane-08/09 published PRINT_ENABLE while the firmware was still asking for
its shared surfaces and never got an acknowledgment, where rg-ane-04..06
with the same kernel did: a doorbell the firmware misses while it
initializes is not repeated. While a command's slot still belongs to the
firmware, ring its doorbell again every 10 ms; count the re-rings and log
them, and on a startup timeout log the IO slot header.

Assisted-by: LLM
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AGCaKaz9NWJpjHQ9RwzeMj
…orkspace ane)

Job commands record the system counter (get_cycles, 24 MHz on arm64, the
clock of the firmware's event timestamps) at publish and at the slot's
completion, events at consumption, as u64 trailers counted by the record's
trailer word. h16_postboot=1 does, after startup, the three engine writes
macOS 25F84 ends ANE_Init with (0x1880058=1, 0x1880060=0, 0x196c000|=1)
and logs the values before and after.

Assisted-by: LLM
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AGCaKaz9NWJpjHQ9RwzeMj
…(diagnostic, workspace ane)

An EVENT step remembers terminal events of other UUIDs it consumes, so a job
can publish several EXECUTEs before waiting for any. The transport never
reuses pool space; a job wraps it back to its first command once no EXECUTE
is outstanding, so long timing jobs do not run out of pool.

Assisted-by: LLM
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AGCaKaz9NWJpjHQ9RwzeMj
astaff added 26 commits November 1, 2026 15:39
@astaff

astaff commented Nov 3, 2026

Copy link
Copy Markdown
Contributor Author

Replaced by #920, a smaller series on the merged H16 driver. The diagnostics here were written before #720 merged and served their purpose.

@astaff astaff closed this Nov 3, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant