Re: [linux-sunxi] [PATCH] media: cedrus: Implement runtime PM

[Ezequiel Garcia <ezequiel@vanguardiasur.com.ar>]

Hi Jernej, Paul:

On Mon, 20 Apr 2020 at 13:41, Jernej Škrabec <jernej.skrabec@gmail.com> wrote:
>
> Dne ponedeljek, 20. april 2020 ob 17:10:10 CEST je Paul Kocialkowski
> napisal(a):
> > Hi,
> >
> > On Sun 19 Apr 20, 15:28, Samuel Holland wrote:
> > > On 4/8/20 11:01 AM, Jernej Škrabec wrote:
> > > > Hi Samuel!
> > > >
> > > > Dne sreda, 08. april 2020 ob 03:02:32 CEST je Samuel Holland napisal(a):
> > > >> This allows the VE clocks and PLL_VE to be disabled most of the time.
> > > >>
> > > >> Since the device is stateless, each frame gets a separate runtime PM
> > > >> reference. Enable autosuspend so the PM callbacks are not run before
> > > >> and
> > > >> after every frame.
> > > >>
> > > >> Signed-off-by: Samuel Holland <samuel@sholland.org>
> > > >> ---
> > > >>
> > > >> I tested this with v4l2-request-test. I don't have the setup to do
> > > >> anything more complicated at the moment.
> > > >>
> > > >> ---
> > > >>
> > > >>  drivers/staging/media/sunxi/cedrus/cedrus.c   |   7 ++
> > > >>  .../staging/media/sunxi/cedrus/cedrus_hw.c    | 115 ++++++++++++------
> > > >>  .../staging/media/sunxi/cedrus/cedrus_hw.h    |   3 +
> > > >>  3 files changed, 88 insertions(+), 37 deletions(-)
> > >
> > > [snip]
> > >
> > > > Reset above causes problem. When format is set in cedrus_s_fmt_vid_cap()
> > > > a
> > > > function is called, which sets few registers in HW. Of course, there is
> > > > no
> > > > guarantee that someone will start decoding immediately after capture
> > > > format is set. So, if the driver puts VPU to sleep in the mean time,
> > > > reset will clear those registers and decoded video will be in different
> > > > format than expected. It could be even argued that registers should not
> > > > be set in that function and that this is design issue or bug in driver.
> > >
> > > You're right. I didn't see that cedrus_dst_format_set() was called outside
> > > cedrus_engine_enable/disable().
> >
> > This might indeed be an issue with multiple decoding contexts in parallel,
> > with potentially different formats. For that reason, it looks like the
> > right thing to do would be to set the format at each decoding run based on
> > the format set in the context by s_fmt.
>
> So you are suggesting that cedrus_dst_format_set() should be moved to
> cedrus_device_run(), right? This way all registers are set at each run, which
> is then truly stateless.
>

BTW, this is how the Hantro and Rockchip's Rkvdec
drivers are implemented. One of the main reasons is
to have simultaneous decoding support.

Thanks,
Ezequiel