Re: [PATCH v2 for 5.4 2/4] media: hantro: Fix H264 max frmsize supported on RK3288

[Tomasz Figa]

On Mon, Oct 14, 2019 at 7:10 AM Nicolas Dufresne
<nicolas.dufresne@xxxxxxxxxxxxx> wrote:
>
> Le jeudi 10 octobre 2019 Ã 16:23 +0900, Tomasz Figa a Ãcrit :
> > On Tue, Oct 8, 2019 at 11:12 PM Jonas Karlman <jonas@xxxxxxxxx> wrote:
> > > On 2019-10-08 15:53, Tomasz Figa wrote:
> > > > On Tue, Oct 8, 2019 at 10:35 PM Tomasz Figa <tfiga@xxxxxxxxxxxx> wrote:
> > > > > On Tue, Oct 8, 2019 at 7:42 PM Tomasz Figa <tfiga@xxxxxxxxxxxx> wrote:
> > > > > > On Tue, Oct 8, 2019 at 3:31 PM Jonas Karlman <jonas@xxxxxxxxx> wrote:
> > > > > > > On 2019-10-08 07:27, Tomasz Figa wrote:
> > > > > > > > Hi Ezequiel, Jonas,
> > > > > > > >
> > > > > > > > On Tue, Oct 8, 2019 at 2:46 AM Ezequiel Garcia <ezequiel@xxxxxxxxxxxxx> wrote:
> > > > > > > > > From: Jonas Karlman <jonas@xxxxxxxxx>
> > > > > > > > >
> > > > > > > > > TRM specify supported image size 48x48 to 4096x2304 at step size 16 pixels,
> > > > > > > > > change frmsize max_width/max_height to match TRM.
> > > > > > > > >
> > > > > > > > > Fixes: 760327930e10 ("media: hantro: Enable H264 decoding on rk3288")
> > > > > > > > > Signed-off-by: Jonas Karlman <jonas@xxxxxxxxx>
> > > > > > > > > ---
> > > > > > > > > v2:
> > > > > > > > > * No changes.
> > > > > > > > >
> > > > > > > > >  drivers/staging/media/hantro/rk3288_vpu_hw.c | 4 ++--
> > > > > > > > >  1 file changed, 2 insertions(+), 2 deletions(-)
> > > > > > > > >
> > > > > > > > > diff --git a/drivers/staging/media/hantro/rk3288_vpu_hw.c b/drivers/staging/media/hantro/rk3288_vpu_hw.c
> > > > > > > > > index 6bfcc47d1e58..ebb017b8a334 100644
> > > > > > > > > --- a/drivers/staging/media/hantro/rk3288_vpu_hw.c
> > > > > > > > > +++ b/drivers/staging/media/hantro/rk3288_vpu_hw.c
> > > > > > > > > @@ -67,10 +67,10 @@ static const struct hantro_fmt rk3288_vpu_dec_fmts[] = {
> > > > > > > > >                 .max_depth = 2,
> > > > > > > > >                 .frmsize = {
> > > > > > > > >                         .min_width = 48,
> > > > > > > > > -                       .max_width = 3840,
> > > > > > > > > +                       .max_width = 4096,
> > > > > > > > >                         .step_width = H264_MB_DIM,
> > > > > > > > >                         .min_height = 48,
> > > > > > > > > -                       .max_height = 2160,
> > > > > > > > > +                       .max_height = 2304,
> > > > > > > > This doesn't match the datasheet I have, which is RK3288 Datasheet Rev
> > > > > > > > 1.4 and which has the values as in current code. What's the one you
> > > > > > > > got the values from?
> > > > > > > The RK3288 TRM vcodec chapter from [1], unknown revision and date, lists 48x48 to 4096x2304 step size 16 pixels under 25.5.1 H.264 decoder.
> > > > > > >
> > > > > > > I can also confirm that one of my test samples (PUPPIES BATH IN 4K) is 4096x2304 and can be decoded after this patch.
> > > > > > > However the decoding speed is not optimal at 400Mhz, if I recall correctly you need to set the VPU1 clock to 600Mhz for 4K decoding on RK3288.
> > > > > > >
> > > > > > > I am not sure if I should include a v2 of this patch in my v2 series, as-is this patch do not apply on master (H264_MB_DIM has changed to MB_DIM in master).
> > > > > > >
> > > > > > > [1] http://www.t-firefly.com/download/firefly-rk3288/docs/TRM/rk3288-chapter-25-video-encoder-decoder-unit-(vcodec).pdf
> > > > > > I checked the RK3288 TRM V1.1 too and it refers to 3840x2160@24fps as
> > > > > > the maximum.
> > > > > >
> > > > > > As for performance, we've actually been getting around 33 fps at 400
> > > > > > MHz with 3840x2160 on our devices (the old RK3288 Asus Chromebook
> > > > > > Flip).
> > > > > >
> > > > > > I guess we might want to check that with Hantro.
> > > > > Could you check the value of bits 10:0 in register at 0x0c8? That
> > > > > should be the maximum supported stream width in the units of 16
> > > > > pixels.
> > > > Correction: The unit is 1 pixel and there are additional 2 most
> > > > significant bits at 0x0d8, 15:14.
> > >
> > > I will check this later tonight when I have access to my devices.
> > > The PUPPIES BATH IN 4K (4096x2304) sample decoded without issue using rockchip 4.4 BSP kernel and mpp last time I tested.
> > >
> > > The vcodec driver in 4.4 BSP kernel use 300/400 Mhz as default clock rate and will change to 600 Mhz when width is over 2560, see [1]:
> > >   raise frequency for resolution larger than 1440p avc
> > >
> > > [1] https://github.com/rockchip-linux/kernel/blob/develop-4.4/drivers/video/rockchip/vcodec/vcodec_service.c#L2551-L2570
> >
> > How comes it works for us well at 400 MHz? Better DRAM? Differences in
> > how Vcodec BSP handles the hardware that somehow make the decoding
> > slower?
>
> FWIW, here on the mainline driver, on RK3288, playing a 4K30 sample
> (probably the max for this one) get stuck at 20fps with 400MHz. So
> 600MHz would in theory be perfect to reach 30fps. That being said,
> different stream yield different performance with H264 and other
> CODECs, so doing a completely objective evaluation is hard.

For a fair comparison, we're using the following stream in our 4K
performance test:
http://storage.googleapis.com/chromiumos-test-assets-public/tast/cros/video/perf/h264/2160p_30fps_300frames_20190801.h264

Best regards,
Tomasz