вт, 7 жовт. 2025 р. о 22:37 Luca Ceresoli <[email protected]> пише:
>
> Hello Svyatoslav,
>
> On Tue Oct 7, 2025 at 6:02 PM CEST, Svyatoslav Ryhel wrote:
> > пн, 6 жовт. 2025 р. о 21:55 Luca Ceresoli <[email protected]> пише:
> >>
> >> Hello Svyatoslav,
> >>
> >> On Thu Oct 2, 2025 at 8:20 AM CEST, Svyatoslav Ryhel wrote:
> >> >> > > > 12 represents amount of bits used per pixel, 8 for Y plane, 2 for 
> >> >> > > > U
> >> >> > > > plane and 2 for V plane, total is 12. "but explainable with a 
> >> >> > > > comment
> >> >> > > > and improve-able later" why then we cannot use 12 with a comment? 
> >> >> > > > this
> >> >> > > > is all arbitrary. Downstream is not wrong from this perspective, 
> >> >> > > > you
> >> >> > > > don't take into account that YUV420 is planar and it uses 3 
> >> >> > > > planes a
> >> >> > > > whole Y plane and 1/4 of U and V which in total results in wigth 
> >> >> > > > + 2 *
> >> >> > > > 1/4 width which is width * 3/2
> >> >> > >
> >> >> > > Yes -- but AIUI, the only thing the bpp value is used for the 
> >> >> > > bytesperline calculation. When we add the special case for planar 
> >> >> > > formats, which doesn't use the bpp value, then the value 12 is 
> >> >> > > never used anywhere. We should at least have a comment saying it is 
> >> >> > > unused. (At that point, we could just hardcode the bpp values in 
> >> >> > > the fmt_align function -- but I don't mind either way.)
> >> >> > >
> >> >> > https://ffmpeg.org/pipermail/ffmpeg-user/2023-June/056488.html
> >> >>
> >> >> I understand very well that for YUV420, each pixel has 12 bits of color 
> >> >> information. But how many bits of color information each pixel has is 
> >> >> not useful in the context of this driver. The number of bytes per line 
> >> >> is not related to how many bits of color information each pixel has for 
> >> >> planar formats.
> >> >
> >> > No, it has direct impact. This is how buffer size / image size is
> >> > calculated since we place each plane consecutive. And bytes per line
> >> > is used specifically in image size calculation. This is common part
> >> > with non-planar formats. Then since Tegra provides a dedicated
> >> > channels/buffers for each plane, configuration of planar format
> >> > includes an additional step with calculation for each plane.
> >>
> >> Sorry, I haven't followed the discussion in detail, but I tested you series
> >> on Tegra20 VIP and capture does not work, with a SIGSEGV in
> >> gstreamer. Bisecting pointed to this as the first commit where the issue
> >> happens.
> >>
> >> I compared the input and output values of tegra20_fmt_align() at this
> >> commit and at the previous one, and this is the result:
> >>
> >>                        before this patch     with this patch
> >>   At function entry:
> >>   bpp                        1                     12
> >>   pix->width                 640                   640
> >>   pix->height                480                   480
> >>
> >>   On return:
> >>   pix->bytesperline          640                   960
> >>   pix->sizeimage             460800                460800
> >>
> >> I hope these info will help.
> >
> > Which command did you use? I have tested with ffmpeg and
> > yuv422/yuv420p and it worked perfectly fine.
>
> I have a simple testing script that runs these commands, with
> VNODE="/dev/video0":
>
> v4l2-ctl -d ${VNODE} --set-ctrl horizontal_flip=1 --set-ctrl vertical_flip=1
>
> gst-launch-1.0 -ve v4l2src device=${VNODE} num-buffers=500 \
>   ! video/x-raw,width=640,height=480,framerate=50/1,format=I420 \
>   ! videorate drop-only=true skip-to-first=true \
>   ! video/x-raw,framerate=50/4 \
>   ! queue \
>   ! avenc_mpeg4 \
>   ! mp4mux \
>   ! filesink location=/tmp/grab.mp4
>
> Luca

I can reproduce what you are observing. ok, I will drop this commit.

> --
> Luca Ceresoli, Bootlin
> Embedded Linux and Kernel engineering
> https://bootlin.com

Reply via email to