Skip to content

Pixel 8 Pro GPU: zuma's CMU_G3D, ACPM GPU clock, panthor and a GPU thermal zone - #157

Closed
greenbrettmichael wants to merge 12 commits into
derekross:omamobilefrom
greenbrettmichael:omamobile-zuma-husky-gpu
Closed

greenbrettmichael wants to merge 12 commits into
derekross:omamobilefrom
greenbrettmichael:omamobile-zuma-husky-gpu

Conversation

@greenbrettmichael

@greenbrettmichael greenbrettmichael commented Oct 31, 2026 •

Copy link
Copy Markdown
Contributor

12 commits on omamobile, rebased after #153 landed. Review updates are in the comments: the TMU entry next to #169's gs201, the G3D CMU_TOP clock ids at 22-34 to leave 16-21 to #163, the #87 wording, and a comment that keeps 940 MHz the last row of zumapro's table.

This turns on the Pixel 8 Pro's GPU, a Mali-G715 MC7, with panthor, DVFS to stock's 890 MHz and a GPU thermal zone. It follows the Pixel 9a's path (#84, #85, #86, #87 and W4HO's panthor match) closely, because zuma's GPU is the same generation and the integration matches almost everywhere.

What zuma shares with zumapro (checked against Google's zuma tree and husky's stock device tree):

  • CMU_G3D at 0x1ee00000, with the same offsets, field widths and parents in Google's zuma CMUCAL tables as zumapro's, PLL0822x M/P/S layout included. The same goes for CMU_TOP's G3D GLB, NOCD and SWITCH paths.
  • The ACPM DVFS domains: zumapro's 17, in the same order (Google's enum acpm_dvfs_id is zuma's), so G3D is 7 and G3DL2 is 8, as stock's gpu1_cmu_cal_id/gpu0_cmu_cal_id name them. DVFS is on channel 0, as stock's cal_if node has it.
  • The GPU DVFS table: husky's stock gpu_dvfs_table_v2 is the 9a's pair for pair, but stock stops at 890 MHz (gpu_dvfs_max_freq), so zuma's paired clock reuses zumapro's table without its 940 MHz level.
  • The ASN hash: stock's l2-hash-values = <0x36db6d 0x5b6db6 0x00> is zumapro's (W4HO's commit already cites Google's zuma-gpu.dtsi for those values), so google,zuma-mali takes zumapro's panthor data.
  • The TMU: stock reads the SoC temperatures through ACPM channel 9, with zumapro's seven sensors in the same order (BIG, MID, LITTLE, G3D, ISP, TPU, AUR), so gs101_acpm_tmu takes zumapro's data.
  • The node: 0x1f000000, size 0x480000, JOB/MMU/GPU on SPI 347/348/346, the same as zumapro's. There's no power domain: stock's own PD driver reports pd-g3d: on at probe, so the bootloader leaves G3D powered.

The commits:

  1. dt-bindings: clock: google: Add the Tensor G3 G3D CMU
  2. clk: samsung: zuma: Add CMU_G3D: read-only, as on zumapro.
  3. dt-bindings: firmware: google,gs101-acpm-ipc: Add the Tensor G3 clock ids: google,zuma-acpm.h, plus its MAINTAINERS line.
  4. clk: samsung: acpm: Add Tensor G3 (zuma) with a paired GPU clock: zuma_gpu_max_khz, 0 (read-only) by default.
  5. dt-bindings: gpu: mali-valhall-csf: Add the Tensor G3 GPU
  6. drm/panthor: Add Tensor G3 cache hash configuration
  7. arm64: dts: exynos: google: zuma: add CMU_G3D and the GPU (both disabled)
  8. arm64: dts: exynos: google: zuma-husky: enable the GPU: core on ACPM's paired gpu clock, and OPPs 150–649 MHz.
  9. dt-bindings: firmware: google,gs101-acpm-ipc: Add the Tensor G3 TMU
  10. thermal: samsung: gs101_acpm_tmu: Add the Tensor G3
  11. arm64: dts: exynos: google: zuma: add the ACPM TMU and a GPU thermal zone: it throttles panthor's devfreq from 85 C and is critical at 105 C, as arm64: dts: exynos: google: zumapro: CPU DVFS, thermal zones and the GPU's top levels #87 proposes for zumapro, below stock's own GPU trips (passive from 100 C, hot at 115 C).
  12. arm64: dts: exynos: google: zuma-husky: Add the GPU's levels to 890 MHz: stock's top.

Tested on husky (tags husky-test-20261031-g1, the first 8 commits, and -g2, all 12, each merged with #156 for Wi-Fi):

  • The clocks: CMU_G3D reads back the bootloader's level: the G3D PLL at 149.9 MHz and the L2 PLL at 159.744 MHz, with the four GPU gates on. clk-acpm binds zuma-acpm-clk, and all 17 domains read plausible rates from the firmware: g3d 150 MHz, g3dl2 160 MHz, gpu 150 MHz.
  • panthor: Mali-G715 id 0xb8a2 major 0x1 minor 0x2, shader_present=0x1110055 (7 cores). The CSF firmware (arm/mali/arch11.8/mali_csffw.bin, linux-firmware) boots with interface v2.8.0, and /dev/dri/renderD128 appears.
  • Rendering: Mesa 26.2.4 reports Mali-G715 MC7 (Panfrost), OpenGL ES 3.1. A surfaceless EGL/GLES2 test renders a triangle into an FBO and reads back the expected pixels, with no GL errors and no GPU faults.
  • DVFS: with zuma_gpu_max_khz=649000, devfreq moves 150 → 649 MHz under load, and the firmware reads back g3d 649 MHz with g3dl2 749 MHz, the paired level. It returns to 150/160 at idle. The draw rate doubles. At the default (0), the GPU stays at the bootloader's 150 MHz.
  • Thermal: gs101_acpm_tmu binds and registers gpu-thermal, which reads 52 C idle, and panthor's devfreq becomes its 13-level cooling device.
    • Sustained load: with zuma_gpu_max_khz=890000, a heavy fragment shader on a 2048x2048 FBO held 890 MHz (g3dl2 900 MHz) for 3 minutes. It ran at 11 frames/s, against 2 at 150 MHz. The GPU leveled off at 82-83 C, under the 85 C trip, with no faults.
    • Throttling: with the trip lowered to 65 C through sysfs, step_wise took the GPU down 850, 807, 723 to 649 MHz (cooling states 1-4) and held it there at about 72 C, then let it idle once the load stopped. The trip was then set back to 85 C.
  • The rest of the phone: touch and Wi-Fi are unaffected.

panthor logs _opp_set_regulators: no regulator (mali) found: voltage stays with the ACPM firmware, as on gs201 and zumapro.

tools/kernel-pr-check --build --merge passes on the first 8 commits: it merges cleanly onto 51fb40d, checkpatch --strict has nothing to read, W=1 builds of the 6 touched C files pass, 8 device trees have no new dtbs_check warnings, and dt_binding_check passes on 8 bindings. The thermal commits pass the same checks on top.

Also seen, and not here yet:

🤖 Generated with Claude Code

@derekross

Copy link
Copy Markdown
Owner

Thanks Brett, this is a careful port. I checked the DVFS ids, the 890 MHz cap, the hash values and the CPU limits against Google's zuma tree. The Pixel 6, 7 and 9a device trees come out byte-identical, and clk-acpm, panthor and the TMU driver are unchanged for them function by function. Three small things: please push the g1, g2 and c1 test tags (I can't find them on your fork), and paste a uname -r and a tools/phone-smoke report from the g2 boot. The commits say "as on zumapro" for the trips, but those zones are still in open #87, so please say "as #87 proposes". Optional: one decoded read of the firmware's hardware thermal trip after TMU INIT. This merges after #153.

@derekross

Copy link
Copy Markdown
Owner

Thanks Brett. #153 is in now (rebase-merged, so its commits have new ids), and #169 added the Tensor G2 to the same TMU binding and driver, so this needs a rebase onto omamobile. The TMU conflict is small: keep the gs201 entry and add zuma's next to it, and say "gs201, zuma and zumapro" in the comment and the binding.

One more thing I want to flag clearly: Lef's #163 also adds CMU_TOP clock ids starting at 16 in google,zuma.h, and git merges your header and his without a conflict, leaving duplicate ids. I plan to merge #163 first, so your G3D TOP ids would move to 22-34 (and TOP's nr_clk_ids to CLK_DOUT_G3D_SWITCH + 1). I've built that resolution and every commit compiles; I can push it to a branch for you if that's easier.

The round-1 points are still open too: please push the g1, g2 and c1 tags (I still can't find them on your fork), paste uname -r and a tools/phone-smoke report from the g2 boot, and make the trip wording point at #87. I checked the Pixel 7 Pro side on the rebased tree: its device tree is byte-identical, and clk-acpm, panthor and the TMU driver compile to the same instructions for it. Non-blocking: nr_pairs = ARRAY_SIZE(zumapro_gpu_pairs) - 1 depends on 940 MHz being the last row; a comment or BUILD_BUG_ON would protect it.

@greenbrettmichael

Copy link
Copy Markdown
Contributor Author

Thanks Derek. All done and force-pushed:

  • Rebased onto omamobile. The TMU conflict is resolved as you suggested: gs201's entry stays and zuma's goes next to it. The binding and the driver comment say "gs201, zuma and zumapro", and the Kconfig title says "Tensor G2, G3 and G4".
  • The G3D CMU_TOP ids are 22-34 now, and TOP's nr_clk_ids is still CLK_DOUT_G3D_SWITCH + 1. I test-merged Pixel 8 Pro battery: CMU_PERIC1, USI11 and the MAX77759 bus #163 onto the branch. The header merges with no duplicate ids, and clk-zuma.c and the binding conflict visibly instead, so whichever lands second gets a real conflict. Thanks for the offer of the branch; it was quick to do here.
  • The tags are on my fork now: husky-test-20261031-g1, -g2 and -c1. Also -d1 to -d3, -w11 and -w12 for the other PRs, and today's husky-test-20261101-e1. Sorry, I'd only pushed the older ones.
  • The trip wording: the commit says "as arm64: dts: exynos: google: zumapro: CPU DVFS, thermal zones and the GPU's top levels #87 proposes for zumapro". The device-tree comments no longer compare with zumapro.
  • 940 MHz: a comment on the 940 MHz row ("keep this level last") and on zuma's nr_pairs. I didn't use BUILD_BUG_ON, because a static const table's contents aren't a constant expression in C.
  • The hardware trip read (the optional one) isn't done yet.

tools/kernel-pr-check --build --merge passes on the 12 commits, plus #158's one. It merges cleanly onto caf5f8a2b, W=1 builds of the 4 touched C files pass, there are no new dtbs_check warnings on 8 trees, and dt_binding_check passes on 3 bindings.

The boot: I no longer have the g2 boot, so the report is from today's e1. That's this branch as pushed, merged with #158, #160 and #156 (717fba4f, rebased), with modules-push, slot B. panthor renders the desktop, and the hottest zone reads 50 °C at idle.

tools/phone-smoke on husky-test-20261101-e1

uname -r: 7.3.0-rc6-omamobile-husky-wifi (the test kernels pin the release name; this is build #11 of tag husky-test-20261101-e1).

Check Result Detail
boot (required) ok kernel 7.3.0-rc6-omamobile-husky-wifi, slot b, up 289s
storage guard (required) ok 62 partitions; only userdata and slot B's boot images writable
root (required) ok ext4 on userdata, 97G free
kernel health (required) ok no oops or BUG; 0 WARN, 0 error lines in dmesg
systemd units ok none failed
USB gadget (required) ok bound to 11210000.usb
display ok card2-DSI-1 connected, 1344x2992
GPU ok render node on panthor
session (required) ok Hyprland and the Omarchy shell running
lock ok PAM file present, the shell can lock (passwordPam true, locked true)
touch ok SALIXP Touchscreen
buttons ok gpio-keys
Wi-Fi ok wlp1s0 on brcmfmac, link up
Bluetooth skip no controller
battery skip no battery power supply
sound cards skip none
sensors skip no sensor proxy
cameras skip no libcamera tools
thermal ok hottest zone 50 °C

The "0 error lines" is only because I cleared the ring buffer with dmesg -C during the screen-off test on this boot. From journalctl -k -b -p err, the boot's error lines, leaving out brcmfmac's, are:

  • panthor: _opp_set_regulators: no regulator (mali) found, expected: ACPM sets the rail;
  • the HK3's one refused brightness write before allow_tx (sending dcs data 51 02 bc failed: -1), whose level is kept and sent again;
  • exynos-clkout: cannot match parent device, from the zuma PMU already on omamobile, and on the d3 boot too;
  • systemd's gpt-auto generator, and "unable to open an initial console".

@greenbrettmichael greenbrettmichael changed the title Pixel 8 Pro GPU: zuma's CMU_G3D, ACPM GPU clock, panthor and a GPU thermal zone (stacked on #153) Pixel 8 Pro GPU: zuma's CMU_G3D, ACPM GPU clock, panthor and a GPU thermal zone Nov 1, 2026
@derekross

Copy link
Copy Markdown
Owner

Thanks Brett, everything I asked for is in:

CI is green here, with dtbs_check and dt_binding_check. Since my last look, clk-acpm.c and the TMU driver changed only in comments and one match entry, so what I checked for the Pixel 7 Pro still holds: its device tree doesn't change, and its code paths are the same.

The merge plan: Lef's #163 goes first, then your display PR #160, which merges cleanly on top of it. After that, #158 (which carries these 12 commits) needs one rebase over both, with conflicts in clk-zuma.c, the clock binding, zuma.dtsi and zuma-husky.dts. Once it's rebased and green, I'll merge #158 and close this one as merged with it. Nothing more is needed here; the hardware trip read stays optional.

@derekross

Copy link
Copy Markdown
Owner

Brett, for when you're back: Somnius's #200 adds CMU_TOP IDs 22-27 in the same block of google,zuma.h as your 22-34, and git merges the two without a conflict, so we'd end up with duplicate IDs. #200 goes first. When you rebase #157 and #158, could you start your CMU_TOP IDs at 28, with nr_clk_ids to match? Thank you!

Document zuma's (Tensor G3) CMU_G3D, which clocks the GPU, and the
CMU_TOP clocks it takes: the G3D GLB, NOCD and SWITCH paths and the
shared PLL dividers behind them. The inputs and their order are those of
zumapro's CMU_G3D, which has the same layout.

Assisted-by: LLM
Signed-off-by: Brett-Michael Green <brett@bmgvisualtech.com>
Describe the GPU's clocks on zuma (Tensor G3): CMU_TOP's G3D GLB, NOCD
and SWITCH paths, with the shared PLL dividers they select from, and
CMU_G3D at 0x1ee00000, with its two PLLs, muxes, dividers and the GPU's
core, coregroup and stacks gates.

As on zumapro, firmware sets these up and keeps G3D on, and the GPU's
rate changes go through ACPM, so everything here is read-only. Google's
zuma CMUCAL tables give the same offsets, field widths and parents as
zumapro's CMU_G3D: the PLLs have PLL0822x's M/P/S layout, and the
L2_GLB and STACKS dividers have no ratio field.

Link: https://android.googlesource.com/kernel/google-modules/soc/gs/+/refs/heads/android-gs-shusky-6.1-android16/drivers/soc/google/cal-if/zuma/cmucal-sfr.c
Assisted-by: LLM
Signed-off-by: Brett-Michael Green <brett@bmgvisualtech.com>
… ids

The Tensor G3 (zuma) ACPM firmware numbers its DVFS domains as Google's
enum acpm_dvfs_id for zuma: MIF, INT, CPUCL0-2, DSU, BCI, G3D, G3DL2,
TPU, INTCAM, TNR, CAM, MFC, DISP, AOC and BW, the same 17 as the
Tensor G4's. Add their clock ids, and the paired GPU clock after them
(17), as the gs201 and zumapro headers do, so device trees can name
them.

Link: https://android.googlesource.com/kernel/google-modules/soc/gs/+/refs/heads/android-gs-shusky-6.1-android16/drivers/soc/google/cal-if/zuma/acpm_dvfs_zuma.h
Assisted-by: LLM
Signed-off-by: Brett-Michael Green <brett@bmgvisualtech.com>
The ACPM protocol driver creates zuma-acpm-clk for the Tensor G3, but
clk-acpm had no match for it. Add the zuma variant, built like the
zumapro one:

- the DVFS domains as clocks. zuma's are zumapro's 17, in the same
  order (Google's enum acpm_dvfs_id is zuma's), so G3D is 7 and G3DL2
  is 8, as the Pixel 8 Pro's stock GPU node names them
  (gpu1_cmu_cal_id 0xb040007, gpu0_cmu_cal_id 0xb040008);
- a paired gpu clock (17) that moves G3D and G3DL2 together. The
  Pixel 8 Pro's stock gpu_dvfs_table_v2 is the Pixel 9a's, pair for
  pair, so zumapro's table is reused, without its 940 MHz level: zuma's
  gpu_dvfs_max_freq is 890 MHz;
- zuma_gpu_max_khz, 0 (read-only) by default, as on gs201 and zumapro.

The firmware's DVFS channel is 0, as stock's cal_if node has it.

Assisted-by: LLM
Signed-off-by: Brett-Michael Green <brett@bmgvisualtech.com>
The Tensor G3 (zuma) has a Mali-G715, arch 11.8.10 r1p2 on the
Pixel 8 Pro, like the Tensor G4's. Add its compatible, so that panthor
can match its integration-specific cache hash, as it does zumapro's.

Assisted-by: LLM
Signed-off-by: Brett-Michael Green <brett@bmgvisualtech.com>
The Tensor G3 (zuma) integrates the same Mali-G715 as the Tensor G4,
with the same ASN hash values: the Pixel 8 Pro's stock device tree has
l2-hash-values = <0x36db6d 0x5b6db6 0x00>, from Google's zuma-gpu.dtsi,
the source zumapro's match already cites. On zumapro, CSF startup fails
with a GPU bus fault without them.

Match google,zuma-mali with zumapro's SoC data.

Link: https://android.googlesource.com/kernel/devices/google/zuma/+/42395e8adcb880aa336f0da6bb2a19ed53929c40/dts/zuma-gpu.dtsi
Assisted-by: LLM
Signed-off-by: Brett-Michael Green <brett@bmgvisualtech.com>
Describe CMU_G3D and the Mali-G715 at 0x1f000000, both disabled, as
zumapro does: the GPU's registers, its JOB, MMU and GPU interrupts
(SPI 347, 348 and 346) and its size come from the Pixel 8 Pro's stock
mali node, and its core, coregroup and stacks clocks from CMU_G3D.

There is no power domain: stock's own power-domain driver reports
pd-g3d on at probe, so the bootloader leaves G3D powered.

Assisted-by: LLM
Signed-off-by: Brett-Michael Green <brett@bmgvisualtech.com>
Turn on CMU_G3D and the Mali-G715 on the Pixel 8 Pro:

- core is ACPM's paired gpu clock (ZUMA_CLK_ACPM_GPU), which sets G3D
  and G3DL2 together through the GPU DVFS pairs;
- coregroup and stacks stay on CMU_G3D's gates;
- an OPP table with the shader rates 150-649 MHz.

Stock goes up to 890 MHz. The levels from 723 MHz on are left out until
a GPU thermal zone is in the device tree, as on the Pixel 9a, because
nothing upstream throttles the GPU yet.

Assisted-by: LLM
Signed-off-by: Brett-Michael Green <brett@bmgvisualtech.com>
The Tensor G3 (zuma) reads its SoC temperatures through the same ACPM
TMU service as gs101 and zumapro, with zumapro's seven sensors in the
same order: the Pixel 8 Pro's stock device tree names them BIG, MID,
LITTLE, G3D, ISP, TPU and AUR (ids 0-6) on ACPM channel 9. Add its
compatible.

Assisted-by: LLM
Signed-off-by: Brett-Michael Green <brett@bmgvisualtech.com>
The Tensor G3 (zuma) TMU service is the Tensor G4's: ACPM channel 9 and
seven sensors in the same order, BIG, MID, LITTLE, G3D, ISP, TPU and
AUR, as the Pixel 8 Pro's stock device tree has them (acpm_tmu on
channel 9, tmuctrl_0 to tmuctrl_6). Match google,zuma-acpm-tmu with
zumapro's data.

Assisted-by: LLM
Signed-off-by: Brett-Michael Green <brett@bmgvisualtech.com>
…zone

Add zuma's TMU sensor under acpm_ipc and a gpu-thermal zone on its G3D
sensor (3). It throttles panthor's devfreq from 85 C and shuts down at
105 C, as derekross#87 proposes for zumapro, below the stock Pixel 8 Pro's own
GPU trips (passive from 100 C, hot at 115 C). For that, the GPU node
gets #cooling-cells.

Assisted-by: LLM
Signed-off-by: Brett-Michael Green <brett@bmgvisualtech.com>
With gpu-thermal throttling panthor's devfreq, add the Pixel 8 Pro's
GPU levels from 723 to 890 MHz, stock's top (gpu_dvfs_max_freq).

Assisted-by: LLM
Signed-off-by: Brett-Michael Green <brett@bmgvisualtech.com>
@greenbrettmichael

Copy link
Copy Markdown
Contributor Author

Thanks Derek. Done: the CMU_TOP G3D ids start at 28 now (28-40), with nr_clk_ids to match. This branch has the same 12 commits as #158's rebased head, on today's omamobile. Details and the gate are on #158.

@derekross

Copy link
Copy Markdown
Owner

Thanks, Brett. The ids are right: 28-40 after #200's, with no duplicates in the header. As
planned, these twelve commits land with #158, so I'll close this one with a link when #158 merges;
my notes on the stack are over there.

@derekross

Copy link
Copy Markdown
Owner

Landed with #158, which carried these commits.

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.

2 participants