lkml.org 
[lkml]   [2022]   [Mar]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH v1 22/24] media: hantro: h264: Make dpb entry management more robust
On Wed, Mar 30, 2022 at 12:16 PM Nicolas Dufresne
<nicolas.dufresne@collabora.com> wrote:
>
> Le mercredi 30 mars 2022 à 09:59 +0200, Sebastian Fricke a écrit :
> > Hey Nicolas,
> >
> > On 28.03.2022 15:59, Nicolas Dufresne wrote:
> > > From: Jonas Karlman <jonas@kwiboo.se>
> > >
> > > The driver maintains stable slot location for reference pictures. This
> >
> > s/slot location/slot locations/
> >
> > > change makes the code more robust by using the reference_ts as key and
> > > by marking all entries invalid right from the start.
> > >
> > > Signed-off-by: Jonas Karlman <jonas@kwiboo.se>
> > > Signed-off-by: Nicolas Dufresne <nicolas.dufresne@collabora.com>
> > > ---
> > > drivers/staging/media/hantro/hantro_h264.c | 10 ++++------
> > > 1 file changed, 4 insertions(+), 6 deletions(-)
> > >
> > > diff --git a/drivers/staging/media/hantro/hantro_h264.c b/drivers/staging/media/hantro/hantro_h264.c
> > > index 228629fb3cdf..7377fc26f780 100644
> > > --- a/drivers/staging/media/hantro/hantro_h264.c
> > > +++ b/drivers/staging/media/hantro/hantro_h264.c
> > > @@ -258,8 +258,7 @@ static void prepare_table(struct hantro_ctx *ctx)
> > > static bool dpb_entry_match(const struct v4l2_h264_dpb_entry *a,
> > > const struct v4l2_h264_dpb_entry *b)
> > > {
> > > - return a->top_field_order_cnt == b->top_field_order_cnt &&
> > > - a->bottom_field_order_cnt == b->bottom_field_order_cnt;
> > > + return a->reference_ts == b->reference_ts;
> > > }
> > >
> > > static void update_dpb(struct hantro_ctx *ctx)
> > > @@ -273,13 +272,13 @@ static void update_dpb(struct hantro_ctx *ctx)
> > >
> > > /* Disable all entries by default. */
> > > for (i = 0; i < ARRAY_SIZE(ctx->h264_dec.dpb); i++)
> > > - ctx->h264_dec.dpb[i].flags &= ~V4L2_H264_DPB_ENTRY_FLAG_ACTIVE;
> > > + ctx->h264_dec.dpb[i].flags = 0;
> >
> > Ehm ... we just remove all flags? Can't this have any unwanted side
> > effects like removing a flag that we actually wanted to keep?
> > (Like long term or the field flags?)
>
> This is a local copy of the dpb, the DPB is fully passed for every decode. So
> these flags will be fully restored lower when we copy the found entry. In fact,
> holding a state here would not represent well the userland intention and can
> have negative side effect on the decoding. Flags are not immutable between
> decode and can change. This simplify the code, and make things less error prone.
> This part of the code is already a bit complex, no need for an extra layer.
>
> > If we just want to set the DPB entry inactive, then removing the ACTIVE
> > flag seems like the correct approach to me.
> > If we want to get rid of the VALID flag as well, then we could just do:
> > ctx->h264_dec.dpb[i].flags &=
> > ~(V4L2_H264_DPB_ENTRY_FLAG_ACTIVE | V4L2_H264_DPB_ENTRY_FLAG_VALID);
> >
> > In case we really want to reset all flags, I'd say adjust the comment
> > above it:
> > ```
> > - /* Disable all entries by default. */
> > + /* Reset the flags for all entries by default. */
> > ```
>
> This reads the same to me, but I can do that yes. understand that VALID means
> the reference exist and the TS should point to some existing past reference
> (unless there was some decode error, which the userland may not be aware yet as
> this is asynchronous). While ACTIVE means that it is used as a reference. FFMPEG
> is known not to filter inactive references. ACTIVE is just a flag without bunch
> of other flags that can change for every decode. So none of this make sense
> between 2 decodes.
>

Please leave the comment as-is, 'Disable all entries' is readable.

As much as I'd love to have very clear comments everywhere,
we have to accept that only people with insight on codec details
are expected to make sense of driver details :-)

Thanks,
Ezequiel

\
 
 \ /
  Last update: 2022-03-31 01:44    [W:0.205 / U:0.092 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site