drm/rockchip: vop2: ACLK_VOP pinned at 500 MHz starves DP 4K120 on RK3588

[Igor Paunovic]

Hi Chaoyi,

That is exactly what I was missing, thank you. Having the vendor's
decision flow meant I could stop guessing and go measure the case where
my rule and yours disagree. I have done that now, and the result is
worth more than my earlier argument was.

First, on equivalence: I do not think my threshold is equivalent to the
linedur_ns term - it is strictly narrower. With
linedur_ns = crtc_htotal * 1000000 / crtc_clock, and htotal 4400 for the
4K modes my panel offers:

  3840x2160 at 120  1188000 kHz   linedur_ns 3703
  3840x2160 at 60    594000 kHz   linedur_ns 7407
  2560x1440 at 144  ~586000 kHz   linedur_ns ~4600-4700

My VOP2_HIGH_BW_PIXCLK_KHZ of 1000000, at htotal 4400, is the same as
saying linedur_ns < 4400 - about 1.7x stricter than your 7500. So my
rule declines to boost in cases where yours boosts, and 3840x2160 at 60 is
one of them: your rule raises it on two counts (crtc_hdisplay > 2560 and
linedur_ns 7407 < 7500), mine does not raise it at all.

So I measured that case. I unplugged both HDMI cables, leaving the
DisplayPort output as the only display, and switched it between
3840x2160 at 120 and 3840x2160 at 60, reading aclk_vop and dclk_vop2 from
clk_summary at each point:

                        4K120 (YCbCr 4:2:0)      4K60
  dclk_vop2             594 MHz                  594 MHz
  aclk_vop              750 MHz                  500 MHz
  vop interrupts/s      120                      60
  POST_BUF_EMPTY        0                        0, over 60 s

The dclk is identical in the two states, because the 4K120 link runs
YCbCr 4:2:0 and the 4K60 link does not. The interface rate did not move
at all. What moved was what the video port composes - a full-screen 4K
ARGB plane at 120 Hz versus the same plane at 60 Hz - and the AXI clock.
That is the separation I told you I could not make from my earlier data,
where mode and composition moved together. The interrupt counts are one
vblank per frame in both states, so there is no starvation at either
point.

3840x2160 at 60 is therefore clean at ACLK 500 MHz with a full-screen 4K
plane. For that case the downstream rule boosts and does not need to.

Putting all four points I now have in terms of composed pixels per ACLK
cycle:

  3840x2160 at 120   995.3 Mpx/s   500 MHz   1.99   storms (116630 irq/s)
  3840x2160 at 120   995.3 Mpx/s   750 MHz   1.33   clean
  2560x1440 at 144   530.8 Mpx/s   500 MHz   1.06   clean
  3840x2160 at 60    497.7 Mpx/s   500 MHz   1.00   clean

The boundary sits between 1.33 and 1.99. That quantity does not depend
on the connector or on the pixel format, which is what I think you were
pointing me at, and it fits every point I have rather than the two I
started from.

The honest limit of this: all four are a single full-screen plane on a
single video port. They say nothing about plane_num_4k or crtc_num > 1,
which are exactly the terms my condition has no counterpart for. I am
not proposing to drop your terms - I am saying the two mode-derived ones
appear to carry margin for this shape of workload, and I would rather
ask than assume.

The other thing I had not appreciated is where the decision lives.
rockchip_drm_aclk_adjust() runs from atomic_commit_tail, so it is
re-evaluated on every commit with that commit's plane information, and
aclk_adjust_frame_num holds the boost for two more commits on the way
down. My unsent patch puts the decision in vop2_crtc_atomic_enable() and
_disable(), which only run on modeset. A condition that depends on what
the planes are composing cannot live there: a client swapping a 1080p
plane for a 4K one without a modeset would never be seen. So the
placement has to change, not just the condition.

Two things I would still rather learn than guess:

  - Is the two-commit hold on the way down a hardware requirement - the
    rate has to be up before the frame that needs it and stay up for a
    beat after - or is it belt and braces? It decides whether an
    upstream version needs the same hysteresis.

  - Are the mode-derived terms describing something the hardware stalls
    on, or are they a conservative stand-in for "this is probably a
    heavy scene"? My 4K60 and 1440p144 points only make sense to me
    under the second reading.

I will rewrite the patch along the lines of your flow rather than the
pixel-clock threshold, and move the decision to the commit path. For
what it is worth, the version I have been running locally since 7 August
has driven 3840x2160 at 120 HDR over DisplayPort with no HDMI output
present at all, which is the case the FRL-gated condition cannot serve;
it also drops the rate back to 500 MHz on modeset and restores it, so
the refcount side behaves.

Thanks again - this turned a guess into a measurement.

Igor