`AMDGPU_PP_SENSOR_GPU_AVG_POWER` is not supported on some GPUs/APUs

The AMDGPU driver now internally converts AMDGPU_INFO_SENSOR_GPU_AVG_POWER to AMDGPU_PP_SENSOR_GPU_AVG_POWER.
However, AMDGPU_PP_SENSOR_GPU_AVG_POWER is not supported on some GPUs/APUs (SMU7, Vega10, Vega12, SMU12 [Renoir], Yellow Carp, SMU13.0.{5,6}).
Those AMD GPUs support only AMDGPU_PP_SENSOR_GPU_INPUT_POWER.

agd5f/linux@9366c2e8
agd5f/linux@47f1724d

With this change, applications that were getting AMD GPU power consumption from the libdrm API will receive an error.
https://github.com/Syllo/nvtop/issues/244

@superm1

This should be expected behavior now. The values being returned previously were wrong for average power. Those products can only return instantaneous power.

changed the description

So, are there any plans to add something like AMDGPU_INFO_SENSOR_GPU_INPUT_POWER to libdrm and support it in the AMDGPU driver?
Currently, applications have no way to get SMU7, Vega10, and Vega12 power consumption information from libdrm .

These patches should fix it.

0002-drm-amdgpu-fall-back-to-INPUT-power-for-AVG-power-vi.patch

0003-drm-amdgpu-add-new-INFO-IOCTL-query-for-input-power.patch

0001-drm-amdgpu-fix-avg-vs-input-power-reporting-on-smu7.patch

If mesa wants to expose both input and avg power, it'll need to use the new query.

@emansom
Can you try these patches?

I tested patch 2/3 and patch 3/3 on Cezanne APU (support only AMDGPU_PP_SENSOR_GPU_INPUT_POWER).

Test branch of libdrm (libdrm-amdgpu-sys-rs): https://github.com/Umio-Yasuno/libdrm-amdgpu-sys-rs/tree/input_power

Looks good to me.

@Umio-Yasuno Apologies, forgot about this. Testing this on a patched 6.6-cachyos-rc7. Can confirm it works as expected on this Polaris RX 550:

(See the POW in nvtop in the bottom right corner now showing the input power usage once again, instead of the N/A shown before)

(If you are wondering why the clocks are so low and the wattage so low: i'm running with a custom PowerPlay Profile to optimize for power efficiency)

What about with/without for amdgpu_top? Does that require this patch too?

amdgpu_top changed to get GFX {Average,Input} power from hwmon since v0.2.3, so it is not affected by these patch.

https://github.com/Umio-Yasuno/amdgpu_top/commit/b520bfcf613784187f375fff22fc7caea7106e55
https://github.com/Umio-Yasuno/amdgpu_top/commit/98efbf0b3803261dc87c9c5e8c1e89c2103731a0

changed title from AMDGPU_PP_SENSOR_GPU_AVG_POWER is not supported in SMU7, Vega10, Vega12 to AMDGPU_PP_SENSOR_GPU_AVG_POWER is not supported in some GPUs/APUs

changed the description

changed title from AMDGPU_PP_SENSOR_GPU_AVG_POWER is not supported in some GPUs/APUs to AMDGPU_PP_SENSOR_GPU_AVG_POWER is not supported on some GPUs/APUs

changed the description

mentioned in issue mesa/mesa#10100 (closed)

This bug breaks viewing iGPU power consumption on my system using https://flathub.org/apps/io.missioncenter.MissionCenter (which uses nvtop under the hood).

Are there any blocks to getting the fix patches merged and backported to stable kernels? Changes to kernel APIs breaking userspace apps is Not A Good Thing, and in this case is affecting real-world users.

If there's anything I can do to help speed this along, please let me know ASAP.

Just wanted to check in if there has been any progress in getting this regression fix merged.

Patches have been sent to the mailing list, trying to get someone to review them.

mentioned in commit agd5f/linux@3819f50b

mentioned in commit agd5f/linux@54147f1c

mentioned in commit agd5f/linux@d0206985

mentioned in commit nouveau@836e236b

mentioned in commit nouveau@1db180f5

Patches have been merged in 6.7.2, this issue is now resolved for me. Tested on a AMD Radeon Polaris RX 580.

Awesome to hear! Any word on these getting backported to 6.6.x?

closed