Re: [PATCH v4] media: vivid: Improve timestamping

[Arnd Bergmann]

On Mon, Dec 3, 2018 at 10:15 AM Hans Verkuil <hverkuil@xxxxxxxxx> wrote:
>
> On 12/02/2018 09:43 PM, Arnd Bergmann wrote:
> > On Sun, Dec 2, 2018 at 2:47 PM Gabriel Francisco Mandaji
> > <gfmandaji@xxxxxxxxx> wrote:
> >
> >> @@ -667,10 +653,28 @@ static void vivid_overlay(struct vivid_dev *dev, struct vivid_buffer *buf)
> >>         }
> >>  }
> >>
> >> +static void vivid_cap_update_frame_period(struct vivid_dev *dev)
> >> +{
> >> +       u64 f_period;
> >> +
> >> +       f_period = (u64)dev->timeperframe_vid_cap.numerator * 1000000000;
> >> +       do_div(f_period, dev->timeperframe_vid_cap.denominator);
> >> +       if (dev->field_cap == V4L2_FIELD_ALTERNATE)
> >> +               do_div(f_period, 2);
> >> +       /*
> >> +        * If "End of Frame", then offset the exposure time by 0.9
> >> +        * of the frame period.
> >> +        */
> >> +       dev->cap_frame_eof_offset = f_period * 9;
> >> +       do_div(dev->cap_frame_eof_offset, 10);
> >> +       dev->cap_frame_period = f_period;
> >> +}
> >
> > Doing two or three do_div() operations is going to make this rather
> > expensive on 32-bit architectures, and it looks like this happens for
> > each frame?
> >
> > Since each one is a multiplication followed by a division, could this
> > be changed to using a different factor followed by a bit shift?
>
> The division by 2 can obviously be replaced by a shift, and the
> 'End of Frame' calculation can be simplified as well by multiplying by
> 7 and dividing by 8 (again a simple shift): this equals 0.875 which is
> close enough to 0.9 (so update the comment as well).

The first division

      f_period = (u64)dev->timeperframe_vid_cap.numerator * 1000000000;
      do_div(f_period, dev->timeperframe_vid_cap.denominator);

looks like it could be replaced with a fixed multiplication with a
precomputed 'u64 factor = div_u64(numerator * 100000000, denominator)

> It's all a bit overkill since this function isn't called very often,
> but these are easy changes to make.

Ah, I assumed it was called once per frame or more. If this is only done
during initalization, it doesn't matter.

        Arnd