Skip to content

omamobile-device-tegu 0.2.23: the GPU's lowest level is 419 MHz, not 150 MHz - #922

Draft
W4HO-ham wants to merge 1 commit into
derekross:mainfrom
W4HO-ham:tegu-gpu-floor
Draft

W4HO-ham wants to merge 1 commit into
derekross:mainfrom
W4HO-ham:tegu-gpu-floor

Conversation

@W4HO-ham

@W4HO-ham W4HO-ham commented Nov 19, 2026 •

Copy link
Copy Markdown
Contributor

Draft, from reading this again: the evidence is thinner than the text below says.

  • The boot check used a local rule with ACTION=="add|change", not this one (add).
  • On my daily kernel the GPU's levels come from a bring-up unit that binds the GPU again, which is not how the published kernel does it. On the published kernel (omamobile at 26867fc4e) the floor read 419000000 at the start of 40 boots last night, with that local rule; I did not record the levels or the cooling device there.
  • min_freq's store returns -EAGAIN until devfreq has added its QoS request, just after the device is registered, so an add rule can lose a race. It would fail safe, at 150 MHz.
  • "Higher isn't better" rests on one run at 649 MHz; nothing between 150 and 419 MHz was tried.
  • The frame times were read from dri/1/state a thousand times a second, which takes the mode-setting locks the commits need: fine for comparing two floors, weaker as absolute counts.

Owed before this is ready: boots with this exact rule on the published kernel and the daily one (min_freq, available_frequencies, the udev journal, the cooling device), the DEVICES GPU cell brought up to date, and a smoke report.

On the Pixel 9a the GPU's devfreq governor (simple_ondemand) looks at the load every 50 ms and starts from 150 MHz, so a swipe or an app opening after a quiet moment began at 150 MHz and its first frames came late while the governor caught up. This sets the floor to 419 MHz, with one udev rule line that fires when panthor's devfreq device appears (from the graphics start).

Frames. On my palmtop (Pixel 9a, kernel 7.3.0-rc6-omamobile-tegu-s2mpg15-m1-g991042e2eff3-dirty, Hyprland 0.57.0 with Omarchy's animations on), I timed each frame the compositor commits, by the display plane's framebuffer changing in /sys/kernel/debug/dri/1/state, read every millisecond. A block is 12 transitions: two each of a home swipe over a terminal, the terminal opening and the terminal closing. Frames later than 12.5 ms (one and a half periods at 120 Hz), with the floor at 150 and 419 MHz in turn:

150 MHz 419 MHz
first pair of blocks 38 17
second pair 41 20

An earlier run of 9 transitions each gave 31 at 150 MHz, 14 at 419 MHz and 28 at 649 MHz, so higher isn't better. Raising the CPU floors, or keeping the compositor's and the shell's threads off the little cores, changed nothing.

Cost. Idle, panthor suspends the GPU: on a lit home screen it was awake 538 ms in a minute. Battery drain, cable in with the charger's buck off (CHG_CNFG_00 0x00 under a dead-man), screen at 500 of 3628, three rounds of 60 s means, in mA:

150 MHz 419 MHz
lit home screen, idle 619, 575, 597 602, 576, 595
UI loop (open a terminal, swipe home, close it) 924, 1036, 956 919, 956, 973

No difference above the noise either way.

At boot: the same rule on my palmtop set min_freq to 419000000 once the GPU's devfreq device appeared.

Tests: test/all checks the installed rule and that its one line sets the 9a GPU's floor and nothing else; it passes, as does tools/pr-check origin/main. The version is 0.2.23 because #912 is 0.2.22; either can go first, and the second gets a one-line rebase.

…150 MHz

simple_ondemand looks at the GPU's load every 50 ms, so a swipe or an app opening after a quiet moment started at 150 MHz and its first frames were late while the governor caught up. A udev rule sets the devfreq floor to 419 MHz when panthor's device appears.

On a Pixel 9a with animations on, over 12 transitions each (home swipes, a terminal opening and closing), frames later than 12.5 ms fell from 38 and 41 to 17 and 20. Idle, panthor suspends the GPU (538 ms awake in a minute on a lit home screen), and the battery drain didn't change: 597 against 591 mA idle and 972 against 949 mA in a UI loop, cable in with the charger off.
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