[DjDeveloperr]

On the MacBook Neo (A18 Pro, G17P), agx_stats reported only utilisation: pstate, achieved_frequency_hz and busy_ns read unsupported, and fdinfo had no frequency keys. The A18 Pro's 26A428 firmware sends the same statistics records that #1611 decodes on the M5. Until now only the M5 decoded them.

What changes

• g17::telemetry::Protocol describes what a firmware sends beyond utilisation: its highest performance state (tags 0x19/0x1b) and its utilisation period (tag 0x13 timestamps). g17::stats::protocol() picks it from the init graph:
  • 25G83 (M4 Pro/Max): nothing beyond utilisation;
  • A18 Pro: states 1-10, 16 ms period;
  • M5: states 1-13, 8 ms period.
• A18 Pro pstate and busy_ns use the decoder and integration already in place for the M5.
• achieved_frequency_hz and drm-curfreq-*/drm-maxfreq-* take the state's frequency from a per-state table. On the M5 it is the board's table. On the A18 Pro it is the device tree operating points that its performance table A names.
  • The records don't name the table. But at the top state, where the GPU runs while work is in flight (the driver's full effort is a floor), a sustained FP32 FMA load finishes faster than table B's top frequency allows.
  • gpubench compute.heavy kind=alu, 2.147 GFLOP per dispatch, 1.18 ms per dispatch at depth on the J700, needs at least 1.42 GHz on its 5 cores at one FMA per lane per clock. Table B's top state is 1326 MHz; table A's is 1470 MHz.
• Gated cores read state 0 (off), valid, at 0 Hz, in agx_stats and drm-curfreq-*, instead of a stale state. This also changes the M5: it used to read stale while gated. Power-up still starts a new epoch, stale until the next state report.
• Tag 0x10 is no longer decoded on a firmware that reports its state in 0x19/0x1b.
• Docs: sysfs-driver-asahi-agx-stats and asahi-usage-stats.rst.

Measured on the J700, from tracepoint captures of the default kernel (asahi_g17_report + asahi_g17_complete, 24 GPU stress traces)

• The utilisation record arrives every 16.00 ms (median of 1269).
• Integrating it as on the M5 gives a busy time within about 20% of the union of the commands' own execution intervals for loads busy more than half the time. Like the M5, it reads low for loads sparser than one period.
• The 0x19 interval records are periodic samples (about 7% of wall time at any load), not busy intervals, so they don't feed busy_ns.
• While work is in flight the state report reads 10.

Testing

• tools/asahi/run-host-tests.py --only telemetry: 13/13. New A18 Pro decoder and reader tests; every mutation is still killed.
• aurora.ko compile and Neo one-shot test boot (agx_stats, fdinfo, nvtop at idle and under load, kernel fault check): pending. The PR stays a draft until then.

cc @erikwb @RyanTheTide: this extends #1611's decoder to the A18 Pro, and changes the M5's gated reading from stale to state 0 (behaviour only).

🤖 Generated with Claude Code