[RFC] accel/rocket: DVFS on RK3588 - a hardware constraint, and some numbers

[Jiaxing Hu]

Hi Nicolas,

> In the DTS, I've removed the assigned clock/freq, to me keeping it
> would be toward forward compatibility, and is fine, but I don't
> always fully grasp all the subtil trap of DT compatibility

A data point from RK3576, where one clock feeds both NPU cores and the
convolution buffer they share.

With no assigned rate the block comes up at 786.432 MHz, and no board
sets vdd_npu_s0, so it runs on whatever the PMIC boots with: 750 mV
here, where Rockchip's own table asks 800 mV of its 800 MHz step. One
core is exact there. Two cores with jobs in flight at the same time
write single words wrong. A pass is 5400 rows of a batched matrix
multiply, each row against the same multiply done one row at a time:

  786 MHz, 750 mV    11 to 25 wrong rows a pass
  594 MHz, 750 mV    0, 0, 0, 0
  786 MHz, 800 mV    0, 0, 0, 0
  786 MHz, 850 mV    0, 0, 0, 0

So the assigned rate is load bearing until something else carries the
rail. If your OPP table carries it, dropping the rate is better than
keeping it. If the driver may also run with no OPP table, which is
what you say it should do, then the DTS still needs a rate that is
correct at the voltage the board boots with.

Does that hold on RK3588, where the three cores share one clock too?

Cheers,
Jiaxing