On Thursday, 4 December 2025 21:21:08 Central European Standard Time Chia-I Wu 
wrote:
> On Wed, Dec 3, 2025 at 6:04 AM Nicolas Frattaroli
> <[email protected]> wrote:
> >
> > Mali GPUs have three registers that indicate which parts of the hardware
> > are powered and active at any moment. These take the form of bitmaps. In
> > the case of SHADER_PWRACTIVE for example, a high bit indicates that the
> > shader core corresponding to that bit index is active. These bitmaps
> > aren't solely contiguous bits, as it's common to have holes in the
> > sequence of shader core indices, and the actual set of which cores are
> > present is defined by the "shader present" register.
> >
> > When the GPU finishes a power state transition, it fires a
> > GPU_IRQ_POWER_CHANGED_ALL interrupt. After such an interrupt is
> > received, the PWRACTIVE registers will likely contain interesting new
> > information.
> I am seeing
> 
>    irq/342-panthor-412     [000] .....   934.526754: gpu_power_active:
> shader_bitmap=0x0 tiler_bitmap=0x0 l2_bitmap=0x0
>    irq/342-panthor-412     [000] .....   936.640356: gpu_power_active:
> shader_bitmap=0x0 tiler_bitmap=0x0 l2_bitmap=0x0
> 
> on a gpu-bound test. It does not look like collecting samples on
> GPU_IRQ_POWER_CHANGED_ALL gives too much info.

On what GPU and SoC is that? If it's MT8196 then I wouldn't be
surprised if it just broke that hardware register, considering
what it did to the SHADER_PRESENT register.

On RK3588 (v10), GPU_IRQ_POWER_CHANGED_ALL reliably fires when
there is new information available in those registers. I haven't
tried on MT8196 (v13) yet because that still doesn't boot with
mainline so testing anything is a pain.

I don't have any v12 or v11 hardware to test with. From what I
understand, there's no open enough platform to do v11 testing on,
just the Pixel 8 and Pixel 9. I could look into the Cix SoC for v12
though some day, but I don't own one at the moment.

> 
> I think they are more useful to be collected periodically, such that
> we know that in the past X seconds, Y out of a total of Z samples
> indicates activities. That's best done in userspace, and panthor's
> role should be to provide an uapi such as
> https://lore.kernel.org/all/[email protected]/.

This wouldn't give you information on the time a power transition has
completed, which is one of the motivations. A periodically collected
PWRACTIVE would just be roughly correlated to how busy the GPU is,
which isn't very useful additional information as the performance
counters themselves are likely a better source of that kind of info.

What I need to do is restrict this to <= v13 in the next revision
however, because v14 reworks this stuff.

Kind regards,
Nicolas Frattaroli


Reply via email to