[libcamera-devel] [PATCH v2 5/5] ipa: ipu3: af: Simplify accumulations of y_items

[Kate Hsuan]

Hi Folks,

On Thu, Mar 24, 2022 at 8:24 PM Laurent Pinchart via libcamera-devel
<libcamera-devel at lists.libcamera.org> wrote:
>
> Hi Kieran,
>
> On Thu, Mar 24, 2022 at 10:52:38AM +0000, Kieran Bingham wrote:
> > Quoting Laurent Pinchart (2022-03-24 01:06:20)
> > > On Wed, Mar 23, 2022 at 01:56:14PM +0000, Kieran Bingham via libcamera-devel wrote:
> > > > Simplify the accumulation of the total and variance with a ternary
> > > > operator.
> > > >
> > > > Signed-off-by: Kieran Bingham <kieran.bingham at ideasonboard.com>
> > > > ---
> > > >
> > > > This is optional really, it's only really a stylistic preference.
> > > >
> > > >
> > > >  src/ipa/ipu3/algorithms/af.cpp | 14 ++++----------
> > > >  1 file changed, 4 insertions(+), 10 deletions(-)
> > > >
> > > > diff --git a/src/ipa/ipu3/algorithms/af.cpp b/src/ipa/ipu3/algorithms/af.cpp
> > > > index ff5e9fb5b3c5..940ed68ea14a 100644
> > > > --- a/src/ipa/ipu3/algorithms/af.cpp
> > > > +++ b/src/ipa/ipu3/algorithms/af.cpp
> > > > @@ -342,20 +342,14 @@ double Af::afEstimateVariance(Span<const y_table_item_t> y_items, bool isY1)
> > > >       double mean;
> > > >       double var_sum = 0;
> > > >
> > > > -     for (auto y : y_items) {
> > > > -             if (isY1)
> > > > -                     total += y.y1_avg;
> > > > -             else
> > > > -                     total += y.y2_avg;
> > > > -     }
> > > > +     for (auto y : y_items)
> > > > +             total += isY1 ? y.y1_avg : y.y2_avg;
> > > >
> > > >       mean = total / y_items.size();
> > > >
> > > >       for (auto y : y_items) {
> > > > -             if (isY1)
> > > > -                     var_sum += pow((y.y1_avg - mean), 2);
> > > > -             else
> > > > -                     var_sum += pow((y.y2_avg - mean), 2);
> > > > +             double avg = isY1 ? y.y1_avg : y.y2_avg;
> > > > +             var_sum += pow((avg - mean), 2);
> > >
> > > You can drop the extra parentheses.
> > >
> > > Reviewed-by: Laurent Pinchart <laurent.pinchart at ideasonboard.com>
> > >
> > > By, this could also be optimized as
> > >
> > >         double sum = 0;
> > >         double sqr_sum = 0;
> > >
> > >         for (auto y : y_items) {
> > >                 double avg = isY1 ? y.y1_avg : y.y2_avg;
>
> This could be a uint16_t btw.
>
> > >
> > >                 sum += avg;
> > >                 sqr_sum += avg * avg;
> > >         }
> > >
> > >         double mean = total / y_items.size();
> > >         return sqr_sum / y_items.size() - mean * mean;
> > >
> > > which should be more efficient, especially with a larger number of
> > > items. The algorithm is subject to numerical instability though,
> > > especially when the variance is small. The two-pass approach is
> > > numerically stable (when the number of items is small).
> >
> > I did wonder if we could get this in a single loop, but I didn't want to
> > break any assumptions that were made on precision or overflows etc, by
> > making fundamental changes to the operation of the code.
> >
> > I presume that when the variance is small, as a double has 15 decimal
> > points I expect we're way into insignificant numbers that don't affect
> > the overall calculation.
>
> It's not about the decimal point, see
> https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance and
> https://en.wikipedia.org/wiki/Catastrophic_cancellation (I like the name
> "catasrophic cancellation" :-)).

Thanks for JM's detailed description of the variance computation
approaches. The y_table length is determined by the grid width and
height. If the maximum values are used, the y_table length will be
32x24=768. (A fixed number and based on the grid configuration).
Typically, the grid size is set to 16x16=256. Also, the variance value
could be very small (for all white and all black images) or large
(complicated patterns). A certain level of precision of the
floating-point number computation is required.

According to the reasons above, I think keeping it simple is good for
now since y_table length is not too long. If it still impacts the
performance, I can perform some survey on it to see which single pass
variance approach is better.


>
> > Kate, JM, what do you think? Any preference on any of the above?
> >
> > > >       }
> > > >
> > > >       return var_sum / static_cast<double>(y_items.size());
>
> --
> Regards,
>
> Laurent Pinchart
>


--
BR,
Kate