[PATCH v6 2/3] ipa: libipa: Add MatrixInterpolator class
Paul Elder
paul.elder at ideasonboard.com
Mon Jun 10 16:30:40 CEST 2024
On Mon, Jun 10, 2024 at 12:32:23PM +0100, Kieran Bingham wrote:
> Quoting Paul Elder (2024-06-07 09:09:05)
> > Add a class to encapsulate the functionality of fetching a matrix based
> > on an integer key, and interpolating if there is no exact match. This is
> > expected to be used by both color correction matrices / crosstalk
> > correction as well as lens shading correction.
> >
> > A cache is included only for exact matches of the key. The caller is
> > expected to decide the tolererance for rounding.
> >
> > Signed-off-by: Paul Elder <paul.elder at ideasonboard.com>
> > Reviewed-by: Stefan Klug <stefan.klug at ideasonboard.com>
> >
> > ---
> > Changes in v6:
> > - fix doxygen
> >
> > Changes in v5:
> > - improve documentation
> > - replace count() with find() != end() (in get())
> > - add cache
> >
> > Changes in v4:
> > - remove stray semicolons
> > - read from the new yaml layout (which mirrors what we have at the
> > moment for lsc), and add keys to make it more flexible
> >
> > Changes in v3:
> > - add a constructor that takes a map of unsigned int -> matrix
> > - s/unit/identity
> > - clear matrices on reset and when reading from yaml
> > - add assert at get()
> >
> > Changes in v2:
> > - initialize to identity matrix
> > - add a function to reset to identity matrix
> > - other minor fixes
> > ---
> > src/ipa/libipa/matrix_interpolator.cpp | 111 +++++++++++++++++++++
> > src/ipa/libipa/matrix_interpolator.h | 131 +++++++++++++++++++++++++
> > src/ipa/libipa/meson.build | 2 +
> > 3 files changed, 244 insertions(+)
> > create mode 100644 src/ipa/libipa/matrix_interpolator.cpp
> > create mode 100644 src/ipa/libipa/matrix_interpolator.h
> >
> > diff --git a/src/ipa/libipa/matrix_interpolator.cpp b/src/ipa/libipa/matrix_interpolator.cpp
> > new file mode 100644
> > index 000000000..064af9b06
> > --- /dev/null
> > +++ b/src/ipa/libipa/matrix_interpolator.cpp
> > @@ -0,0 +1,111 @@
> > +/* SPDX-License-Identifier: BSD-2-Clause */
> > +/*
> > + * Copyright (C) 2019, Raspberry Pi Ltd
> > + * Copyright (C) 2024, Paul Elder <paul.elder at ideasonboard.com>
> > + *
> > + * Helper class for interpolating maps of matrices
> > + */
> > +#include "matrix_interpolator.h"
> > +
> > +#include <algorithm>
> > +#include <string>
> > +
> > +#include <libcamera/base/log.h>
> > +
> > +#include "libcamera/internal/yaml_parser.h"
> > +
> > +#include "matrix.h"
> > +
> > +/**
> > + * \file matrix_interpolator.h
> > + * \brief Helper class for interpolating maps of matrices
> > + */
> > +
> > +namespace libcamera {
> > +
> > +LOG_DEFINE_CATEGORY(MatrixInterpolator)
> > +
> > +namespace ipa {
> > +
> > +/**
> > + * \class MatrixInterpolator
> > + * \brief Class for storing, retrieving, and interpolating matrices
> > + * \tparam T Type of numerical values to be stored in the matrices
> > + * \tparam R Number of rows in the matrices
> > + * \tparam C Number of columns in the matrices
> > + *
> > + * The main use case is to pass a map from color temperatures to corresponding
> > + * matrices (eg. color correction), and then requesting a matrix for a specific
> > + * color temperature. This class will abstract away the interpolation portion.
> > + */
> > +
> > +/**
> > + * \fn MatrixInterpolator::MatrixInterpolator(const std::map<unsigned int, Matrix<T, R, C>> &matrices)
> > + * \brief Construct a matrix interpolator from a map of matrices
> > + * \param matrices Map from which to construct the matrix interpolator
> > + */
> > +
> > +/**
> > + * \fn MatrixInterpolator::reset()
> > + * \brief Reset the matrix interpolator content to a single identity matrix
> > + */
> > +
> > +/**
> > + * \fn int MatrixInterpolator<T, R, C>::readYaml(const libcamera::YamlObject &yaml)
> > + * \brief Initialize an MatrixInterpolator instance from yaml
> > + * \tparam T Type of data stored in the matrices
> > + * \tparam R Number of rows of the matrices
> > + * \tparam C Number of columns of the matrices
> > + * \param[in] yaml The yaml object that contains the map of unsigned integers to matrices
> > + * \param[in] key_name The name of the key in the yaml object
> > + * \param[in] matrix_name The name of the matrix in the yaml object
>
> I don't see key_name, matrix_name in the definition above? (But I do see
> it below in the implementation).
Doxygen seems to match it fine without, so I'll just remove the
parameters...
>
> > + *
> > + * The yaml object is expected to be a list of maps. Each map has two or more
> > + * pairs: one of \a key_name to the key value (usually color temperature), and
> > + * one or more of \a matrix_name to the matrix. This is a bit difficult to
> > + * explain, so here is an example (in python, as it is easier to parse than
> > + * yaml):
> > + * [
> > + * {
> > + * 'ct': 2860,
> > + * 'ccm': [ 2.12089, -0.52461, -0.59629,
> > + * -0.85342, 2.80445, -0.95103,
> > + * -0.26897, -1.14788, 2.41685 ],
> > + * 'offsets': [ 0, 0, 0 ]
> > + * },
> > + *
> > + * {
> > + * 'ct': 2960,
> > + * 'ccm': [ 2.26962, -0.54174, -0.72789,
> > + * -0.77008, 2.60271, -0.83262,
> > + * -0.26036, -1.51254, 2.77289 ],
> > + * 'offsets': [ 0, 0, 0 ]
> > + * },
> > + *
> > + * {
> > + * 'ct': 3603,
> > + * 'ccm': [ 2.18644, -0.66148, -0.52496,
> > + * -0.77828, 2.69474, -0.91645,
> > + * -0.25239, -0.83059, 2.08298 ],
> > + * 'offsets': [ 0, 0, 0 ]
> > + * },
> > + * ]
> > + *
> > + * In this case, \a key_name would be 'ct', and \a matrix_name can be either
> > + * 'ccm' or 'offsets'. This way multiple matrix interpolators can be defined in
> > + * one set of color temperature ranges in the tuning file, and they can be
> > + * retrieved separately with the \a matrix_name parameter.
> > + *
> > + * \return Zero on success, negative error code otherwise
> > + */
> > +
> > +/**
> > + * \fn Matrix<T, R, C> MatrixInterpolator<T, R, C>::get(unsigned int key)
> > + * \brief Retrieve a matrix from the list of matrices, interpolating if necessary
> > + * \param[in] key The unsigned integer key of the matrix to retrieve
> > + * \return The matrix corresponding to the color temperature
> > + */
> > +
> > +} /* namespace ipa */
> > +
> > +} /* namespace libcamera */
> > diff --git a/src/ipa/libipa/matrix_interpolator.h b/src/ipa/libipa/matrix_interpolator.h
> > new file mode 100644
> > index 000000000..27ab532eb
> > --- /dev/null
> > +++ b/src/ipa/libipa/matrix_interpolator.h
> > @@ -0,0 +1,131 @@
> > +/* SPDX-License-Identifier: BSD-2-Clause */
> > +/*
> > + * Copyright (C) 2019, Raspberry Pi Ltd
> > + * Copyright (C) 2024, Paul Elder <paul.elder at ideasonboard.com>
> > + *
> > + * Helper class for interpolating maps of matrices
> > + */
> > +
> > +#pragma once
> > +
> > +#include <algorithm>
> > +#include <map>
> > +#include <string>
> > +#include <tuple>
> > +
> > +#include <libcamera/base/log.h>
> > +
> > +#include "libcamera/internal/yaml_parser.h"
> > +
> > +#include "matrix.h"
> > +
> > +namespace libcamera {
> > +
> > +LOG_DECLARE_CATEGORY(MatrixInterpolator)
> > +
> > +namespace ipa {
> > +
> > +#ifndef __DOXYGEN__
> > +template<typename T, unsigned int R, unsigned int C,
> > + std::enable_if_t<std::is_arithmetic_v<T>> * = nullptr>
> > +#else
> > +template<typename T, unsigned int R, unsigned int C>
> > +#endif /* __DOXYGEN__ */
> > +class MatrixInterpolator
> > +{
> > +public:
> > + MatrixInterpolator()
> > + {
> > + reset();
> > + }
> > +
> > + MatrixInterpolator(const std::map<unsigned int, Matrix<T, R, C>> &matrices)
> > + {
> > + for (const auto &pair : matrices)
> > + matrices_[pair.first] = pair.second;
> > + }
> > +
> > + ~MatrixInterpolator() {}
> > +
> > + void reset()
> > + {
> > + Matrix<T, R, C> identity;
> > + matrices_.clear();
> > + matrices_[0] = identity;
> > + }
> > +
> > + int readYaml(const libcamera::YamlObject &yaml,
> > + const std::string &key_name,
> > + const std::string &matrix_name)
> > + {
> > + int ret;
> > +
> > + matrices_.clear();
> > +
> > + if (!yaml.isList()) {
> > + LOG(MatrixInterpolator, Error) << "yaml object must be a list";
> > + return -EINVAL;
> > + }
> > +
> > + for (const auto &value : yaml.asList()) {
> > + unsigned int ct = std::stoul(value[key_name].get<std::string>(""));
> > + Matrix<T, R, C> matrix;
> > + if ((ret = matrix.readYaml(value[matrix_name])) < 0) {
> > + LOG(MatrixInterpolator, Error) << "Failed to read matrix";
> > + return ret;
> > + }
> > +
> > + matrices_[ct] = matrix;
> > +
> > + LOG(MatrixInterpolator, Debug)
> > + << "Read matrix '" << matrix_name << "' for key '"
> > + << key_name << "' " << ct << ": "
> > + << matrices_[ct].toString();
> > + }
> > +
> > + if (matrices_.size() < 1) {
> > + LOG(MatrixInterpolator, Error) << "Need at least one matrix";
> > + return -EINVAL;
> > + }
> > +
> > + return 0;
> > + }
> > +
> > + Matrix<T, R, C> get(unsigned int ct)
> > + {
> > + ASSERT(matrices_.size() > 0);
> > +
> > + if (matrices_.size() == 1 ||
> > + ct <= matrices_.begin()->first)
> > + return matrices_.begin()->second;
> > +
> > + if (ct >= matrices_.rbegin()->first)
> > + return matrices_.rbegin()->second;
> > +
> > + if (matrices_.find(ct) != matrices_.end())
> > + return matrices_[ct];
> > +
> > + if (cache_.find(ct) != cache_.end())
> > + return cache_[ct];
> > +
> > + /* The above four guarantee that this will succeed */
> > + auto iter = matrices_.upper_bound(ct);
> > + unsigned int ctUpper = iter->first;
> > + unsigned int ctLower = (--iter)->first;
> > +
> > + double lambda = (ct - ctLower) / static_cast<double>(ctUpper - ctLower);
> > + Matrix<T, R, C> ret =
> > + lambda * matrices_[ctUpper] + (1.0 - lambda) * matrices_[ctLower];
> > + cache_[ct] = ret;
>
> Yikes. I like this addition but I'm a little bit worried here.
>
> Imagine some changing light scene causes us to generate 'every' possible
> light colour temperature. This would consume a potentially large amount
> of memory. Particularly if this was interpolating LSC tables.
>
> The cache_ needs to be limited in size ideally as some form of LRU.
>
> I assumed we'd just keep a single cached Matrix originally (which I
> guess is an LRU of size 1...) , but if it's worth adding multiple
> then I think it needs some limits already.
>
>
> As I asked for the cache, I don't mind if it's removed if that's easier
> for the short term and added in on top with an LRU, or changed to store
> only a single cached ct, or an LRU is implemented.
Alright I'll remove it for now.
>
>
> Otherwise, everything else looks fine to me.
Thanks,
Paul
>
>
>
> > + return ret;
> > + }
> > +
> > +private:
> > + std::map<unsigned int, Matrix<T, R, C>> matrices_;
> > +
> > + std::map<unsigned int, Matrix<T, R, C>> cache_;
> > +};
> > +
> > +} /* namespace ipa */
> > +
> > +} /* namespace libcamera */
> > diff --git a/src/ipa/libipa/meson.build b/src/ipa/libipa/meson.build
> > index 2547a5b83..901d78549 100644
> > --- a/src/ipa/libipa/meson.build
> > +++ b/src/ipa/libipa/meson.build
> > @@ -8,6 +8,7 @@ libipa_headers = files([
> > 'fc_queue.h',
> > 'histogram.h',
> > 'matrix.h',
> > + 'matrix_interpolator.h',
> > 'module.h',
> > 'pwl.h',
> > ])
> > @@ -20,6 +21,7 @@ libipa_sources = files([
> > 'fc_queue.cpp',
> > 'histogram.cpp',
> > 'matrix.cpp',
> > + 'matrix_interpolator.cpp',
> > 'module.cpp',
> > 'pwl.cpp',
> > ])
> > --
> > 2.39.2
> >
More information about the libcamera-devel
mailing list