[DjDeveloperr]

Stacked on #270, which adds the GPU activity input this counts submissions through and is itself stacked on #244 → #229.

Mirror of aurora-silicon#563 (commit 264558d, merged there as 50eb52d), cherry-picked with -x.

On the Neo (J700, T8140), light GPU submissions a few milliseconds apart never got the GPU floor (DCS 4, fabric 3). Graphics powers off between them, so they never count as streaming. Each submission then waits on a low fabric, which nothing else holds up while the panel is off. gpubench idle.wake at a 2 ms gap shows it: 192 µs, against 153 µs on macOS.

This gives such work the floor:

• Work with two or more submissions in each of four 4 ms samples in a row (one 60 Hz frame) counts as dense. Frame-paced loads such as video leave a quiet period in every frame and stay below that.
• The floor keeps the PMP back-off and the vote hold that streaming work has.
• A publication doesn't trigger a sample of its own. One taken as each video frame starts finds the CPUs busy and raises the DCS through their vote; that was +0.07 W in video in an earlier version.
• debugfs counts the samples that applied the floor for streaming and for dense work.
• The T8140 GPU node now links its CLPC (apple,clpc = <&clpc>;). Device trees without that link still find the single T8140 CLPC, as a hint the GPU probes without.

Conflicts resolved against omacom's clpc.c

omacom has none of the aurora code below, so it is left out:

• the T6040 host-demand block in the sampler, and the apple,t6041-clpc provider and lookup;
• the debugfs last_vote … reason_bits line, and the reason |= 4/8 bits;
• the floor release/return counters from soc: apple: clpc: count how often the GPU floor is released and comes back #263, which isn't merged. If soc: apple: clpc: count how often the GPU floor is released and comes back #263 lands first, the dense branch should also call apple_clpc_gpu_floor_count(), as it does in aurora; I'll rebase whichever lands second;
• the sampler wake on a cold publication. KVM guest crashes with a "Bad address" error when accessing GPU memory (specifically drm_gem_shmem pfn_mkwrite) on kernel 7.1.5-400.asahi AsahiLinux/linux#563 only stops arming it on T8140, and this tree never had it, so there is nothing to skip. The commit message says so.

The gpu floor samples: streaming … dense … debugfs line sits in the existing gpu: block. The dense logic, the threshold, the T8140 table entry, the DT fallback and the KUnit case are AsahiLinux#563's, unchanged.

Results

Measured on the Neo on the aurora tree, comparing aurora integration d266c70 with and without this change. That tree also has the 100 ms PMP hand-back from #245 and the counters from #263. Boots strictly alternated, ×2 each, 4 rounds of each row per boot. Panel off except the power legs and the "panel on" row.

row	before	after	macOS
idle.wake 2 ms, wake latency	192.5 µs	170.5 µs (−11%)	153
idle.wake 2 ms, panel on	188.1 µs	160.3 µs (−15%)	153
idle.wake 2 ms, awake latency	~170 µs	~148 µs	157.5
idle.wake 20 ms	588.8 µs	588.5 µs	607
compute.light d1	104.6 µs	104.0 µs	–
render fill 1080p d1 / d2 / d4	694 / 575 / 592 µs	694 / 578 / 596 µs	825 / 643 / 653
power: idle / 1080p30 video / browsing	4.91 / 7.05 / 5.00 W	4.88 / 7.06 / 5.00 W	–

• Across the whole idle.wake 2 ms process, including device setup and teardown, the floor is applied in about two thirds of the CLPC samples: roughly 32 dense and 32 streaming out of about 95.
• Video, browsing and idle power are unchanged, and so is the DCS residency in video.
• The awake latency now beats macOS. What's left of the 2 ms gap is the wake cost, about 22 µs, which macOS doesn't show at that gap. That's the firmware's idle power-off, a separate issue.
• No kernel errors on any boot. KUnit apple-clpc-gpu: 4/4 on aurora.
• Regression gate PERFORMANCE-REGRESSION-PASS. Targeted Vulkan CTS: 4687 Pass / 681 NotSupported / 0 Fail. Desktop alive.

The G17 glue (#270) is the same as aurora's at d266c70, so the dense threshold sees the same counts.

Testing on this branch

• Host build with the shipped config + asahi.config: drivers/soc/apple/ and asahi compile, and asahi.o links.
• compile-check.py --per-commit (W=1): pass, 0 new W=1 warnings.
• check-commits.py: signed off, checkpatch clean.
• dt-check.py: pass, 0 new problems; the one affected T8140 DTB validates against the binding from soc: apple: clpc: take GPU activity reports from the G17 driver (stacked on #244) #270.
• KUnit apple-clpc-gpu (arm64 QEMU, 16K pages): 3/3 (apple_clpc_gpu_dense_test plus the two lifetime cases from soc: apple: clpc: take GPU activity reports from the G17 driver (stacked on #244) #270). Aurora's fourth case covers the cold-publication wake, which isn't carried.
• Kernel CI (dispatched): green at 448fc2cfe0f7: run 40799683469. package, KUnit (omacom's aurora.kunitconfig), boot, host tests and the CI self-test pass. A dispatched run skips commits, compile and device trees, so those three were run locally (results above).