Re: [PATCH 03/16] media: rockchip: rga: align stride to 16 bytes

[Sven Püschel]

Hi Nicolas,

On 10/7/25 8:19 PM, Nicolas Dufresne wrote:
> Hi,
>
>
> Le mardi 07 octobre 2025 à 10:31 +0200, Sven Püschel a écrit :
> > Align the stride to a multiple of 16 according to the RGA3 requirements
> > mentioned in the datasheet. This also ensures that the stride of the RGA2
> > is aligned to 4 bytes, as it needs to divide the value by 4 (one word)
> > before storing it in the register.
> >
> > Increasing the stride for the alignment also requires to increase the
> > sizeimage value. This is usually handled by v4l2_fill_pixfmt_mp, but
> > it doesn't allow to set a stride alignment. Therefore use the generated
> > values to calculate the total number of lines to properly update the
> > sizeimage value after the bytesperline has been aligned.
> >
> > Signed-off-by: Sven Püschel <s.pueschel@xxxxxxxxxxxxxx>
> > ---
> >   drivers/media/platform/rockchip/rga/rga.c | 21 +++++++++++++++++++++
> >   1 file changed, 21 insertions(+)
> >
> > diff --git a/drivers/media/platform/rockchip/rga/rga.c
> > b/drivers/media/platform/rockchip/rga/rga.c
> > index
> > 6438119a6c7aeff1e89e7aa95dcd5d2921fefa08..3cb7ce470c47e39d694e8176875a75fad271
> > 7f96 100644
> > --- a/drivers/media/platform/rockchip/rga/rga.c
> > +++ b/drivers/media/platform/rockchip/rga/rga.c
> > @@ -459,6 +459,25 @@ static int vidioc_enum_fmt(struct file *file, void *priv,
> > struct v4l2_fmtdesc *f
> >   	return 0;
> >   }
> >   
> > +static void align_pixfmt(struct v4l2_pix_format_mplane *pix_fmt)
> > +{
> > +	int lines;
> > +	struct v4l2_plane_pix_format *fmt;
> > +
> > +	/*
> > +	 * Align stride to 16 for the RGA3 (based on the datasheet)
> > +	 * To not dismiss the v4l2_fill_pixfmt_mp helper
> > +	 * (and manually write it again), we're approximating the new
> > sizeimage
> > +	 */
> > +	for (fmt = pix_fmt->plane_fmt;
> > +	     fmt < pix_fmt->plane_fmt + pix_fmt->num_planes;
> > +	     fmt++) {
> > +		lines = DIV_ROUND_UP(fmt->sizeimage, fmt->bytesperline);
> > +		fmt->bytesperline = (fmt->bytesperline + 0xf) & ~0xf;
> > +		fmt->sizeimage = fmt->bytesperline * lines;
> Instead of open coding this, describe this with struct v4l2_frmsize_stepwise and
> then use v4l2_apply_frmsize_constraints().
Looking into v4l2_frmsize_stepwise, it only applies to the width/height 
values, whereas I'm interested to control the stride 
byte-alignment/stepping (allowing free range widths/heights).

Do you intent that I (mis)use it like this: 
v4l2_apply_frmsize_constraints(&fmt->bytesperline, &lines, &constraints)  ?

This only replaces one of the three lines in the for loop (where i 
manually do the alignment with & ~0xf) and would also look odd, as i use 
a value in bytes as a width parameter (expected to be in pixels).

Technically the core problem is that the v4l2_fill_pixfmt_mp helper 
doesn't allow to specify a stride byte alignment and just aligns the 
stride based on the format requirements. Therefore I think that maybe 
creating a new v4l2_fill_pixfmt_mp_aligned helper would probably be the 
better solution to get the correct bytesperline value right from the 
start instead of adjusting it afterwards.

Sincerely
    Sven
> Nicolas
>
> > +	}
> > +}
> > +
> >   static int vidioc_g_fmt(struct file *file, void *priv, struct v4l2_format *f)
> >   {
> >   	struct v4l2_pix_format_mplane *pix_fmt = &f->fmt.pix_mp;
> > @@ -474,6 +493,7 @@ static int vidioc_g_fmt(struct file *file, void *priv,
> > struct v4l2_format *f)
> >   		return PTR_ERR(frm);
> >   
> >   	v4l2_fill_pixfmt_mp(pix_fmt, frm->fmt->fourcc, frm->width, frm-
> > > height);
> > +	align_pixfmt(pix_fmt);
> >   
> >   	pix_fmt->field = V4L2_FIELD_NONE;
> >   	pix_fmt->colorspace = frm->colorspace;
> > @@ -496,6 +516,7 @@ static int vidioc_try_fmt(struct file *file, void *priv,
> > struct v4l2_format *f)
> >   				(u32)MIN_HEIGHT, (u32)MAX_HEIGHT);
> >   
> >   	v4l2_fill_pixfmt_mp(pix_fmt, fmt->fourcc, pix_fmt->width, pix_fmt-
> > > height);
> > +	align_pixfmt(pix_fmt);
> >   	pix_fmt->field = V4L2_FIELD_NONE;
> >   
> >   	return 0;