[PATCH v2] ipa: libipa: histogram: Add transform parameter to constructor

[Laurent Pinchart]

On Tue, Apr 23, 2024 at 02:21:20PM +0000, Barnabás Pőcze wrote:
> Hi
> 
> 
> 2024. április 19., péntek 22:44 keltezéssel, Kieran Bingham <kieran.bingham at ideasonboard.com> írta:
> 
> > Quoting Paul Elder (2024-04-19 06:53:36)
> > > Add a parameter to the histogram constructor that takes a transformation
> > > function to apply to all the bins upon construction.
> > >
> > > This is necessary notably for the rkisp1, as the values reported from
> > > the hardware are 20 bits where the upper 16-bits are meaningful integer
> > > values and the lower 4 bits are fractional and meant to be discarded. As
> > > adding a right-shift parameter is probably too specialized, a generic
> > > function is added as a parameter instead.
> > >
> > > Signed-off-by: Paul Elder <paul.elder at ideasonboard.com>
> > >
> > > ---
> > > This used to be "ipa: libipa: histogram: Add rshift parameter to
> > > constructor"
> > >
> > > Changes in v2:
> > > - change rshift parameter to a function parameter
> > > ---
> > >  src/ipa/libipa/histogram.cpp | 6 ++++--
> > >  src/ipa/libipa/histogram.h   | 6 ++++--
> > >  2 files changed, 8 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/src/ipa/libipa/histogram.cpp b/src/ipa/libipa/histogram.cpp
> > > index c1aac59b..2a4095f4 100644
> > > --- a/src/ipa/libipa/histogram.cpp
> > > +++ b/src/ipa/libipa/histogram.cpp
> > > @@ -40,13 +40,15 @@ namespace ipa {
> > >  /**
> > >   * \brief Create a cumulative histogram
> > >   * \param[in] data A pre-sorted histogram to be passed
> > > + * \param[in] transform The transformation function to apply to every bin
> > 
> > I would have likely said "A transformation function to apply to every
> > bin" but ... it doesn't really impact here.
> > 
> > >   */
> > > -Histogram::Histogram(Span<const uint32_t> data)
> > > +Histogram::Histogram(Span<const uint32_t> data,
> > > +                    std::function<uint32_t(const uint32_t)> transform)
> > >  {
> > >         cumulative_.reserve(data.size());
> > >         cumulative_.push_back(0);
> > >         for (const uint32_t &value : data)
> > > -               cumulative_.push_back(cumulative_.back() + value);
> > > +               cumulative_.push_back(cumulative_.back() + transform(value));
> > >  }
> > >
> > >  /**
> > > diff --git a/src/ipa/libipa/histogram.h b/src/ipa/libipa/histogram.h
> > > index 54bb2a19..89a6b550 100644
> > > --- a/src/ipa/libipa/histogram.h
> > > +++ b/src/ipa/libipa/histogram.h
> > > @@ -8,9 +8,9 @@
> > >  #pragma once
> > >
> > >  #include <assert.h>
> > > +#include <functional>
> > >  #include <limits.h>
> > >  #include <stdint.h>
> > > -
> > >  #include <vector>
> > >
> > >  #include <libcamera/base/span.h>
> > > @@ -23,7 +23,9 @@ class Histogram
> > >  {
> > >  public:
> > >         Histogram() { cumulative_.push_back(0); }
> > > -       Histogram(Span<const uint32_t> data);
> > > +       Histogram(Span<const uint32_t> data,
> > > +                 std::function<uint32_t(const uint32_t)> transform =
> > > +                       [](uint32_t x) { return x; });
> > 
> > Well, that default transform is likely quite easy for a compiler to
> > optimise inline so I think this is as generic as we get here!
> 
> Unfortunately, there will be no inlining without LTO here. And std::function has
> not insignificant overhead even if inlined. For inlining to happen reliably and
> without much overhead, the constructor would need to be templated and inline.

Do you mean passing the transform as a template parameter ? That could
be a good idea, I don't expect a big difference in code size.

Paul, could you give it a try, and check how big the constructor gets ?

> > Thanks for reworking.
> > 
> > Reviewed-by: Kieran Bingham <kieran.bingham at ideasonboard.com>
> > 
> > >         size_t bins() const { return cumulative_.size() - 1; }
> > >         uint64_t total() const { return cumulative_[cumulative_.size() - 1]; }
> > >         uint64_t cumulativeFrequency(double bin) const;

-- 
Regards,

Laurent Pinchart