imagebufalgo.h

Go to the documentation of this file.

// Copyright Contributors to the OpenImageIO project.
// SPDX-License-Identifier: Apache-2.0
// https://github.com/AcademySoftwareFoundation/OpenImageIO


// clang-format off

#pragma once
#define OPENIMAGEIO_IMAGEBUFALGO_H

#if defined(_MSC_VER)
// Ignore warnings about DLL exported classes with member variables that are template classes.
// This happens with the std::vector<T> members of PixelStats below.
#  pragma warning (disable : 4251)
#endif

#include <OpenImageIO/imageio.h>
#include <OpenImageIO/imagebuf.h>
#include <OpenImageIO/fmath.h>
#include <OpenImageIO/parallel.h>
#include <OpenImageIO/span.h>
#include <OpenImageIO/vecparam.h>

#include <limits>

#if !defined(__OPENCV_CORE_TYPES_H__) && !defined(OPENCV_CORE_TYPES_H)
struct IplImage;  // Forward declaration; used by Intel Image lib & OpenCV
namespace cv {
    class Mat;
}
#endif



OIIO_NAMESPACE_BEGIN

// forward declarations
class ColorConfig;
class ColorProcessor;
class Filter2D;


/// @defgroup ImageBufAlgo_intro (ImageBufAlgo common principles)
/// @{
///
/// This section explains the general rules common to all ImageBufAlgo
/// functions. Only exceptions to these rules will be explained in the
/// subsequent listings of all the individual ImageBufAlgo functions.
///
///
/// **Return values and error messages**
///
/// Most ImageBufAlgo functions that produce image data come in two forms:
///
/// 1. Return an ImageBuf.
///
///    The return value is a new ImageBuf containing the result image. In
///    this case, an entirely new image will be created to hold the result.
///    In case of error, the result image returned can have any error
///    conditions checked with `has_error()` and `geterror()`.
///
///        // Method 1: Return an image result
///        ImageBuf fg ("fg.exr"), bg ("bg.exr");
///        ImageBuf dst = ImageBufAlgo::over (fg, bg);
///        if (dst.has_error())
///            std::cout << "error: " << dst.geterror() << "\n";
///
/// 2. Pass a destination ImageBuf reference as the first parameter.
///
///    The function is passed a *destination* ImageBuf where the results
///    will be stored, and the return value is a `bool` that is `true` if
///    the function succeeds or `false` if the function fails. Upon failure,
///    the destination ImageBuf (the one that is being altered) will have an
///    error message set.
///
///        // Method 2: Write into an existing image
///        ImageBuf fg ("fg.exr"), bg ("bg.exr");
///        ImageBuf dst;   // will be the output image
///        bool ok = ImageBufAlgo::over (dst, fg, bg);
///        if (! ok)
///            std::cout << "error: " << dst.geterror() << "\n";
///
/// The first option (return an ImageBuf) is a more compact and intuitive
/// notation that is natural for most simple uses. But the second option
/// (pass an ImageBuf& referring to an existing destination) offers
/// additional flexibility, including more careful control over allocations,
/// the ability to partially overwrite regions of an existing image, and the
/// ability for the destination image to also be one of the input images
/// (for example, add(A,A,B) adds B into existing image A, with no third
/// image allocated at all).
///
/// **Region of interest**
///
/// Most ImageBufAlgo functions take an optional ROI parameter that
/// restricts the operation to a range in x, y, z, and channels. The default
/// ROI (also known as `ROI::All()`) means no region restriction -- the
/// whole image will be copied or altered.
///
/// For ImageBufAlgo functions that write into a destination ImageBuf
/// parameter and it is already initialized (i.e. allocated with a
/// particular size and data type), the operation will be performed on the
/// pixels in the destination that overlap the ROI, leaving pixels in the
/// destination which are outside the ROI unaltered.
///
/// For ImageBufAlgo functions that return an ImageBuf directly, or their
/// `dst` parameter that is an uninitialized ImageBuf, the ROI (if set)
/// determines the size of the result image. If the ROI is the default
/// `All`, the result image size will be the union of the pixel data windows
/// of the input images and have a data type determined by the data types of
/// the input images.
///
/// Most ImageBufAlgo functions also respect the `chbegin` and `chend`
/// members of the ROI, thus restricting the channel range on which the
/// operation is performed.  The default ROI constructor sets up the ROI
/// to specify that the operation should be performed on all channels of
/// the input image(s).
///
/// **Constant and per-channel values**
///
/// Many ImageBufAlgo functions take per-channel constant-valued arguments
/// (for example, a fill color). These parameters are passed as
/// `cspan<float>`. These are generally expected to have length equal to the
/// number of channels. But you may also pass a single float which will be
/// used as the value for all channels. (More generally, what is happening
/// is that the last value supplied is replicated for any missing channel.)
///
/// Some ImageBufAlgo functions have parameters of type `Image_or_Const`,
/// which may take either an ImageBuf reference, or a per-channel constant,
/// or a single constant to be used for all channels.
///
/// **Multithreading**
///
/// All ImageBufAlgo functions take an optional `nthreads` parameter that
/// signifies the maximum number of threads to use to parallelize the
/// operation.  The default value for `nthreads` is 0, which signifies that
/// the number of thread should be the OIIO global default set by
/// `OIIO::attribute()`, which itself defaults to be the detected level of
/// hardware concurrency (number of cores available).
///
/// Generally you can ignore this parameter (or pass 0), meaning to use all
/// the cores available in order to perform the computation as quickly as
/// possible.  The main reason to explicitly pass a different number
/// (generally 1) is if the application is multithreaded at a high level,
/// and the thread calling the ImageBufAlgo function just wants to continue
/// doing the computation without spawning additional threads, which might
/// tend to crowd out the other application threads.
///
///@}



/// Image_or_Const: Parameter-passing helper that is a non-owning reference
/// to either an `ImageBuf&`, `ImageBuf*`, per-channel float constant, or a
/// single float constant. This lets us tame the combinatorics of functions
/// where each of several input parameters may be either images or constant
/// values.
class Image_or_Const {
public:
    struct None {};
    Image_or_Const (None) : m_type(NONE) {}
    Image_or_Const (const ImageBuf &img) : m_type(IMG), m_img(&img) {}
    Image_or_Const (const ImageBuf *img) : m_type(IMG), m_img(img) {}
    Image_or_Const (cspan<float> val) : m_type(VAL), m_val(val) {}
    Image_or_Const (const float& val) : m_type(VAL), m_val(val) {}
    Image_or_Const (const std::vector<float>& val) : m_type(VAL), m_val(val) {}
    Image_or_Const (std::initializer_list<const float> val) : m_type(VAL), m_val(val) {}
    Image_or_Const (const float *v, size_t s) : m_type(VAL), m_val(v,s) {}
    Image_or_Const (const float *v, int s) : m_type(VAL), m_val(v,s) {}

    template<size_t N>
    Image_or_Const(const float (&array)[N]) : Image_or_Const(cspan<float>(array)) {}

    bool is_img () const { return m_type == IMG; }
    bool is_val () const { return m_type == VAL; }
    bool is_empty () const { return m_type == NONE; }
    const ImageBuf& img () const { return *m_img; }
    const ImageBuf* imgptr () const { return m_img; }
    cspan<float> val () const { return m_val; }

    void swap (Image_or_Const &other) {
        std::swap (m_type, other.m_type);
        std::swap (m_img, other.m_img);
        std::swap (m_val, other.m_val);
    }
private:
    enum Contents { NONE, VAL, IMG };
    Contents m_type;
    const ImageBuf * m_img = nullptr;
    cspan<float> m_val;
};




namespace ImageBufAlgo {

// old name (DEPRECATED 1.9)
OIIO_DEPRECATED("use parallel_options (1.9)")
typedef parallel_options parallel_image_options;


/// Create an all-black `float` image of size and channels as described by
/// the ROI.
ImageBuf OIIO_API zero (ROI roi, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API zero (ImageBuf &dst, ROI roi={}, int nthreads=0);


/// @defgroup fill (ImageBufAlgo::fill -- fill a region)
/// @{
///
/// Fill an image region with given channel values, either returning a new
/// image or altering the existing `dst` image within the ROI.  Note that the
/// values arrays start with channel 0, even if the ROI indicates that a
/// later channel is the first to be changed.
///
/// Three varieties of fill() exist: (a) a single set of channel values that
/// will apply to the whole ROI, (b) two sets of values that will create a
/// linearly interpolated gradient from top to bottom of the ROI, (c) four
/// sets of values that will be bilinearly interpolated across all four
/// corners of the ROI.

ImageBuf OIIO_API fill (cspan<float> values, ROI roi, int nthreads=0);
ImageBuf OIIO_API fill (cspan<float> top, cspan<float> bottom,
                        ROI roi, int nthreads=0);
ImageBuf OIIO_API fill (cspan<float> topleft, cspan<float> topright,
                        cspan<float> bottomleft, cspan<float> bottomright,
                        ROI roi, int nthreads=0);
bool OIIO_API fill (ImageBuf &dst, cspan<float> values,
                    ROI roi={}, int nthreads=0);
bool OIIO_API fill (ImageBuf &dst, cspan<float> top, cspan<float> bottom,
                    ROI roi={}, int nthreads=0);
bool OIIO_API fill (ImageBuf &dst, cspan<float> topleft, cspan<float> topright,
                    cspan<float> bottomleft, cspan<float> bottomright,
                    ROI roi={}, int nthreads=0);
/// @}


/// Create a checkerboard pattern of size given by `roi`, with origin given
/// by the `offset` values, checker size given by the `width`, `height`,
/// `depth` values, and alternating between `color1[]` and `color2[]`.  The
/// pattern is defined in abstract "image space" independently of the pixel
/// data window of `dst` or the ROI.
ImageBuf OIIO_API checker (int width, int height, int depth,
                           cspan<float> color1, cspan<float> color2,
                           int xoffset, int yoffset, int zoffset,
                           ROI roi, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API checker (ImageBuf &dst, int width, int height, int depth,
                       cspan<float> color1, cspan<float> color2,
                       int xoffset=0, int yoffset=0, int zoffset=0,
                       ROI roi={}, int nthreads=0);


/// Return an image of "noise" in every pixel and channel specified by the
/// roi. There are several noise types to choose from, and each behaves
/// differently and has a different interpretation of the `A` and `B`
/// parameters:
///
/// - "gaussian"   adds Gaussian (normal distribution) noise values with
///                   mean value A and standard deviation B.
/// - "white"      adds independent uniformly distributed values on range
///                [A,B).
/// - "uniform"    (synonym for "white")
/// - "blue"       adds "blue noise" uniformly distributed on range [A,B) but
///                not independent; rather, they are chosen for good spectral
///                properties for sampling and dither.
/// - "salt"       changes to value A a portion of pixels given by B.
///
/// If the `mono` flag is true, a single noise value will be applied to all
/// channels specified by `roi`, but if `mono` is false, a separate noise
/// value will be computed for each channel in the region.
///
/// The random number generator is actually driven by a hash on the "image
/// space" coordinates and channel, independently of the pixel data window
/// of `dst` or the ROI. Choosing different seed values will result in a
/// different pattern, but for the same seed value, the noise at a given
/// pixel coordinate (x,y,z) channel c will is completely deterministic and
/// repeatable.
ImageBuf OIIO_API noise (string_view noisetype,
                         float A = 0.0f, float B = 0.1f, bool mono = false,
                         int seed = 0, ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API noise (ImageBuf &dst, string_view noisetype,
                     float A = 0.0f, float B = 0.1f, bool mono = false,
                     int seed = 0, ROI roi={}, int nthreads=0);


/// Return a const reference to a periodic bluenoise texture with float data
/// in 4 channels that are uncorrelated to each other. Note that unlike most
/// other ImageBufAlgo functions, it does not return an ImageBuf by value, but
/// by const reference.
OIIO_API const ImageBuf& bluenoise_image();


/// Render a single point at (x,y) of the given color "over" the existing
/// image `dst`. If there is no alpha channel, the color will be written
/// unconditionally (as if the alpha is 1.0).
bool OIIO_API render_point (ImageBuf &dst, int x, int y,
                            cspan<float> color=1.0f,
                            ROI roi={}, int nthreads=0);

/// Render a line from pixel (`x1`,`y1`) to (`x2`,`y2`) into `dst`, doing an
/// "over" of the color (if it includes an alpha channel) onto the existing
/// data in `dst`. The `color` should include as many values as
/// `roi.chend-1`. The ROI can be used to limit the pixel area or channels
/// that are modified, and default to the entirety of `dst`. If
/// `skip_first_point` is `true`, the first point (`x1`, `y1`) will not be
/// drawn (this can be helpful when drawing poly-lines, to avoid
/// double-rendering of the vertex positions).
bool OIIO_API render_line (ImageBuf &dst, int x1, int y1, int x2, int y2,
                           cspan<float> color=1.0f,
                           bool skip_first_point = false,
                           ROI roi={}, int nthreads=0);

/// Render a filled or unfilled box with corners at pixels (`x1`,`y1`) and
/// (`x2`,`y2`) into `dst`, doing an "over" of the color (if it includes an
/// alpha channel) onto the existing data in `dst`. The `color` must include
/// as many values as `roi.chend-1`. The ROI can be used to limit the pixel
/// area or channels that are modified, and default to the entirety of
/// `dst`. If `fill` is `true`, the box will be completely filled in,
/// otherwise only its outlien will be drawn.
bool OIIO_API render_box (ImageBuf &dst, int x1, int y1, int x2, int y2,
                          cspan<float> color=1.0f, bool fill = false,
                          ROI roi={}, int nthreads=0);


enum class TextAlignX { Left, Right, Center };
enum class TextAlignY { Baseline, Top, Bottom, Center };

/// Render a text string (encoded as UTF-8) into image `dst`. If the `dst`
/// image is not yet initialized, it will be initialized to be a black
/// background exactly large enough to contain the rasterized text.  If
/// `dst` is already initialized, the text will be rendered into the
/// existing image by essentially doing an "over" of the character into the
/// existing pixel data.
///
/// @param dst
///             Destination ImageBuf -- text is rendered into this image.
/// @param x/y
///             The position to place the text.
/// @param text
///             The text to draw. Linefeed (`\n`) characters are respected
///             as indications that the text spans multiple rows.
/// @param fontsize/fontname
///             Size and name of the font. If the name is not a full
///             pathname to a font file, it will search for a matching font,
///             defaulting to some reasonable system font if not supplied at
///             all), and with a nominal height of fontsize (in pixels).
/// @param textcolor
///             Color for drawing the text, defaulting to opaque white
///             (1.0,1.0,...) in all channels if not supplied. If provided,
///             it is expected to point to a float array of length at least
///             equal to `R.spec().nchannels`, or defaults will be chosen
///             for you).
/// @param alignx/aligny
///             The default behavior is to align the left edge of the
///             character baseline to (`x`,`y`). Optionally, `alignx` and
///             `aligny` can override the alignment behavior, with
///             horizontal alignment choices of TextAlignX::Left, Right, and
///             Center, and vertical alignment choices of Baseline, Top,
///             Bottom, or Center.
/// @param shadow
///             If nonzero, a "drop shadow" of this radius will be used to
///             make the text look more clear by dilating the alpha channel
///             of the composite (makes a black halo around the characters).
///
/// Note that any named fonts (if not a full pathname) will search for the
/// fonts in the following places: (a) any directories named in the global
/// "font_searchpath" attribute or the `$OPENIMAGEIO_FONTS` environment
/// variable; (b) any font-related subdirectories (`fonts`, `Fonts`,
/// `share/fonts`, or `Library/Fonts`) underneath the directories in
/// environment variables `$HOME`, `$SystemRoot`, `$OpenImageIO_ROOT`; (c) a
/// number of common system font areas, including `/usr/share/fonts`,
/// `/Library/fonts`, and `C:/Windows/fonts`; (d) in fonts directories one
/// level up from the place where the currently running binary lives.
bool OIIO_API render_text (ImageBuf &dst, int x, int y, string_view text,
                           int fontsize=16, string_view fontname="",
                           cspan<float> textcolor = 1.0f,
                           TextAlignX alignx = TextAlignX::Left,
                           TextAlignY aligny = TextAlignY::Baseline,
                           int shadow = 0, ROI roi={}, int nthreads=0);


/// The helper function `text_size()` merely computes the dimensions of the
/// text, returning it as an ROI relative to the left side of the baseline
/// of the first character. Only the `x` and `y` dimensions of the ROI will
/// be used. The x dimension runs from left to right, and y runs from top to
/// bottom (image coordinates). For a failure (such as an invalid font
/// name), the ROI will return `false` if you call its `defined()` method.
/// The `text` may contain linefeed characters to designate multiple lines
/// of text.
ROI OIIO_API text_size (string_view text, int fontsize=16,
                        string_view fontname="");


/// Generic channel shuffling: return (or store in `dst`) a copy of `src`,
/// but with channels in the order `channelorder[0..nchannels-1]` (or set to
/// a constant value, designated by `channelorder[i] = -1` and having the
/// fill value in `channelvalues[i]`. In-place operation is allowed (i.e.,
/// `dst` and `src` the same image, but an extra copy will occur).
///
/// @param  nchannels
///             The total number of channels that will be set up in the
///             `dst` image.
/// @param  channelorder
///             For each channel in `dst`, the index of the `src` channel
///             from which to copy. Any `channelorder[i]` < 0 indicates that
///             the channel `i` should be filled with constant value
///             `channelvalues[i]` rather than copy any channel from `src`.
///             If `channelorder` itself is empty, the implied channel order
///             will be `{0, 1, ..., nchannels-1}`, meaning that it's only
///             renaming, truncating, or extending channels, not reordering
///             the channels that are already present.
/// @param  channelvalues Fill values for color channels in which
///             `channelorder[i]` < 0. This can be empty if no channels are
///             to be filled with constant values.
/// @param  newchannelnames
///             An array of new channel names. Channels for which this
///             specifies an empty string will have their name taken from
///             the `src` channel that was copied. If `newchannelnames` is
///             entirely empty, all channel names will simply be copied from
///             `src`.
/// @param  shuffle_channel_names
///             If true, the channel names will be taken from the
///             corresponding channels of the source image -- be careful with
///             this, shuffling both channel ordering and their names could
///             result in no semantic change at all, if you catch the drift.
///             If false (the default), the resulting `dst` image will have
///             default channel names in the usual order ("R", "G", etc.).
///
ImageBuf OIIO_API channels (const ImageBuf &src,
                        int nchannels, cspan<int> channelorder,
                        cspan<float> channelvalues={},
                        cspan<std::string> newchannelnames={},
                        bool shuffle_channel_names=false, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API channels (ImageBuf &dst, const ImageBuf &src,
                        int nchannels, cspan<int> channelorder,
                        cspan<float> channelvalues={},
                        cspan<std::string> newchannelnames={},
                        bool shuffle_channel_names=false, int nthreads=0);


/// Append the channels of `A` and `B` together into `dst` over the region
/// of interest.  If the region passed is uninitialized (the default), it
/// will be interpreted as being the union of the pixel windows of `A` and `B`
/// (and all channels of both images).  If `dst` is not already initialized,
/// it will be resized to be big enough for the region.
ImageBuf OIIO_API channel_append (const ImageBuf &A, const ImageBuf &B,
                                  ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API channel_append (ImageBuf &dst, const ImageBuf &A,
                              const ImageBuf &B, ROI roi={}, int nthreads=0);


/// Return the specified region of pixels of `src` as specified by `roi`
/// (which will default to the whole of `src`, optionally with the pixel
/// type overridden by convert (if it is not `TypeUnknown`).
ImageBuf OIIO_API copy (const ImageBuf &src, TypeDesc convert=TypeUnknown,
                        ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
/// If `dst` is not already initialized, it will be set to the same size as
/// `roi` (defaulting to all of `src`)
bool OIIO_API copy (ImageBuf &dst, const ImageBuf &src, TypeDesc convert=TypeUnknown,
                    ROI roi={}, int nthreads=0);


/// Return the specified region of `src` as an image, without altering its
/// position in the image plane.
///
/// Pixels from `src` which are outside `roi` will not be copied, and new
/// black pixels will be added for regions of `roi` which were outside the
/// data window of `src`.
///
/// Note that the `crop` operation does not actually move the pixels on the
/// image plane or adjust the full/display window; it merely restricts which
/// pixels are copied from `src` to `dst`.  (Note the difference compared to
/// `cut()`).
ImageBuf OIIO_API crop (const ImageBuf &src, ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API crop (ImageBuf &dst, const ImageBuf &src, ROI roi={}, int nthreads=0);


/// Return the designated region of `src`, but repositioned to the image
/// origin and with the full/display window set to exactly cover the new
/// pixel data window. (Note the difference compared to `crop()`).
ImageBuf OIIO_API cut (const ImageBuf &src, ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API cut (ImageBuf &dst, const ImageBuf &src, ROI roi={}, int nthreads=0);


/// Copy `src` pixels within `srcroi` into the `dst` image, offset so that
/// source location (0,0,0) will be copied to destination location
/// (`xbegin`,`ybegin`,`zbegin`). If the `srcroi` is `ROI::All()`, the
/// entirety of the data window of `src` will be used.  It will copy into
/// `channels[chbegin...]`, as many channels as are described by srcroi.
/// Pixels or channels of `src` inside `srcroi` will replace the
/// corresponding destination pixels entirely, whereas `src` pixels outside
/// of `srcroi` will not be copied and the corresponding offset pixels of
/// `dst` will not be altered.
bool OIIO_API paste (ImageBuf &dst, int xbegin, int ybegin,
                     int zbegin, int chbegin, const ImageBuf &src,
                     ROI srcroi={}, int nthreads=0);


/// @defgroup rotateN (rotate in 90 degree increments)
/// @{
///
/// Return (or copy into `dst`) a rotated copy of the image pixels of `src`,
/// in 90 degree increments. Pictorially:
///
///      rotate90             rotate180            rotate270
///     -----------          -----------          -----------
///     AB  -->  CA          AB  -->  DC          AB  -->  BD
///     CD       DB          CD       BA          CD       AC
///

ImageBuf OIIO_API rotate90  (const ImageBuf &src, ROI roi={}, int nthreads=0);
ImageBuf OIIO_API rotate180 (const ImageBuf &src, ROI roi={}, int nthreads=0);
ImageBuf OIIO_API rotate270 (const ImageBuf &src, ROI roi={}, int nthreads=0);
bool OIIO_API rotate90  (ImageBuf &dst, const ImageBuf &src,
                         ROI roi={}, int nthreads=0);
bool OIIO_API rotate180 (ImageBuf &dst, const ImageBuf &src,
                         ROI roi={}, int nthreads=0);
bool OIIO_API rotate270 (ImageBuf &dst, const ImageBuf &src,
                         ROI roi={}, int nthreads=0);
/// @}


/// @defgroup flip-flop-transpose (flip/flop/transpose: mirroring)
/// @{
///
/// Return (or copy into `dst`) a subregion of `src`, but with the scanlines
/// exchanged vertically (flip), or columns exchanged horizontally (flop),
/// or transposed across the diagonal by swapping rows for columns
/// (transpose) within the display/full window. In other words,
///
///        flip                 flop               transpose
///     -----------          -----------          -----------
///     AB  -->  CD          AB  -->  BA          AB  -->  AC
///     CD       AB          CD       DC          CD       BD
///

ImageBuf OIIO_API flip (const ImageBuf &src, ROI roi={}, int nthreads=0);
ImageBuf OIIO_API flop (const ImageBuf &src, ROI roi={}, int nthreads=0);
ImageBuf OIIO_API transpose (const ImageBuf &src, ROI roi={}, int nthreads=0);
bool OIIO_API flip (ImageBuf &dst, const ImageBuf &src,
                    ROI roi={}, int nthreads=0);
bool OIIO_API flop (ImageBuf &dst, const ImageBuf &src,
                    ROI roi={}, int nthreads=0);
bool OIIO_API transpose (ImageBuf &dst, const ImageBuf &src,
                         ROI roi={}, int nthreads=0);
/// @}


/// Return (or store into `dst`) a copy of `src`, but with whatever seties
/// of rotations, flips, or flops are necessary to transform the pixels into
/// the configuration suggested by the "Orientation" metadata of the image
/// (and the "Orientation" metadata is then set to 1, ordinary orientation).
ImageBuf OIIO_API reorient (const ImageBuf &src, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API reorient (ImageBuf &dst, const ImageBuf &src, int nthreads=0);


/// Return a subregion of `src`, but circularly shifting by the given
/// amount.  To clarify, the circular shift of [0,1,2,3,4,5] by +2 is
/// [4,5,0,1,2,3].
ImageBuf OIIO_API circular_shift (const ImageBuf &src,
                                  int xshift, int yshift, int zshift=0,
                                  ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API circular_shift (ImageBuf &dst, const ImageBuf &src,
                              int xshift, int yshift, int zshift=0,
                              ROI roi={}, int nthreads=0);


/// @defgroup rotate (rotate: arbitrary rotation)
/// @{
///
/// Rotate the `src` image by the `angle` (in radians, with positive angles
/// clockwise). When `center_x` and `center_y` are supplied, they denote the
/// center of rotation, in pixel coordinates; in their absence, the rotation
/// will be about the center of the image's display window.
///
/// Only the pixels (and channels) of `dst` that are specified by `roi` will
/// be copied from the rotated `src`; the default `roi` is to alter all the
/// pixels in `dst`. If `dst` is uninitialized, it will be resized to be an
/// ImageBuf large enough to hold the rotated image if recompute_roi is
/// true, or will have the same ROI as `src` if `recompute_roi` is `false`.
/// It is an error to pass both an uninitialized `dst` and an undefined
/// `roi`.
///
/// The filter is used to weight the `src` pixels falling underneath it for
/// each `dst` pixel.  The caller may specify a reconstruction filter by
/// name and width (expressed in pixels units of the `dst` image), or
/// `rotate()` will choose a reasonable default high-quality default filter
/// (lanczos3) if the empty string is passed, and a reasonable filter width
/// if `filterwidth` is 0. (Note that some filter choices only make sense
/// with particular width, in which case this filterwidth parameter may be
/// ignored.)

ImageBuf OIIO_API rotate (const ImageBuf &src, float angle,
                          string_view filtername = string_view(),
                          float filterwidth = 0.0f, bool recompute_roi = false,
                          ROI roi={}, int nthreads=0);
ImageBuf OIIO_API rotate (const ImageBuf &src, float angle,
                          Filter2D *filter, bool recompute_roi = false,
                          ROI roi={}, int nthreads=0);
ImageBuf OIIO_API rotate (const ImageBuf &src,
                          float angle, float center_x, float center_y,
                          string_view filtername = string_view(),
                          float filterwidth = 0.0f, bool recompute_roi = false,
                          ROI roi={}, int nthreads=0);
ImageBuf OIIO_API rotate (const ImageBuf &src,
                          float angle, float center_x, float center_y,
                          Filter2D *filter, bool recompute_roi = false,
                          ROI roi={}, int nthreads=0);
bool OIIO_API rotate (ImageBuf &dst, const ImageBuf &src, float angle,
                      string_view filtername = string_view(),
                      float filterwidth = 0.0f, bool recompute_roi = false,
                      ROI roi={}, int nthreads=0);
bool OIIO_API rotate (ImageBuf &dst, const ImageBuf &src, float angle,
                      Filter2D *filter, bool recompute_roi = false,
                      ROI roi={}, int nthreads=0);
bool OIIO_API rotate (ImageBuf &dst, const ImageBuf &src,
                      float angle, float center_x, float center_y,
                      string_view filtername = string_view(),
                      float filterwidth = 0.0f, bool recompute_roi = false,
                      ROI roi={}, int nthreads=0);
bool OIIO_API rotate (ImageBuf &dst, const ImageBuf &src,
                      float angle, float center_x, float center_y,
                      Filter2D *filter, bool recompute_roi = false,
                      ROI roi={}, int nthreads=0);
/// @}


/// @defgroup resize (resize: resize the image with nicely filtered results)
/// @{
///
/// Set `dst`, over the region of interest, to be a resized version of the
/// corresponding portion of `src` (mapping such that the "full" image
/// window of each correspond to each other, regardless of resolution).
/// If `dst` is not yet initialized, it will be sized according to `roi`.
///
/// The caller may either (a) explicitly pass a reconstruction `filter`, or
/// (b) specify one by `filtername` and `filterwidth`. If `filter` is
/// `nullptr` or if `filtername` is the empty string `resize()` will choose
/// a reasonable high-quality default (blackman-harris when upsizing,
/// lanczos3 when downsizing).  The filter is used to weight the `src`
/// pixels falling underneath it for each `dst` pixel; the filter's size is
/// expressed in pixel units of the `dst` image.

ImageBuf OIIO_API resize (const ImageBuf &src,
                          string_view filtername = "", float filterwidth=0.0f,
                          ROI roi={}, int nthreads=0);
ImageBuf OIIO_API resize (const ImageBuf &src, Filter2D *filter,
                          ROI roi={}, int nthreads=0);
bool OIIO_API resize (ImageBuf &dst, const ImageBuf &src,
                      string_view filtername = "", float filterwidth=0.0f,
                      ROI roi={}, int nthreads=0);
bool OIIO_API resize (ImageBuf &dst, const ImageBuf &src, Filter2D *filter,
                      ROI roi={}, int nthreads=0);
/// @}


/// Set `dst`, over the region of interest, to be a resized version of the
/// corresponding portion of `src` (mapping such that the "full" image
/// window of each correspond to each other, regardless of resolution).  If
/// `dst` is not yet initialized, it will be sized according to `roi`.
/// 
/// Unlike `ImageBufAlgo::resize()`, `resample()` does not take a filter; it
/// just samples either with a bilinear interpolation (if `interpolate` is
/// `true`, the default) or uses the single "closest" pixel (if
/// `interpolate` is `false`).  This makes it a lot faster than a proper
/// `resize()`, though obviously with lower quality (aliasing when
/// downsizing, pixel replication when upsizing).
/// 
/// For "deep" images, this function returns copies the closest source pixel
/// needed, rather than attempting to interpolate deep pixels (regardless of
/// the value of `interpolate`).
///
/// @see ImageBufAlgo::resize()

ImageBuf OIIO_API resample (const ImageBuf &src, bool interpolate = true,
                        ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API resample (ImageBuf &dst, const ImageBuf &src,
                        bool interpolate = true, ROI roi={}, int nthreads=0);


/// @defgroup fit (fit: resize the image with filtering, into a fixed size)
/// @{
///
/// Fit src into `dst` (to a size specified by `roi`, if `dst` is not
/// initialized), resizing but preserving its original aspect ratio. Thus,
/// it will resize to be the largest size with the same aspect ratio that
/// can fit inside the region, but will not stretch to completely fill it in
/// both dimensions.
///
/// The `fillmode` determines which of several methods will be used to
/// determine how the image will fill the new frame, if its aspect ratio
/// does not precisely match the original source aspect ratio:
///     - "width" exactly fills the width of the new frame, either cropping
///       or letterboxing the height if it isn't precisely the right size to
///       preserve the original aspect ratio.
///     - "height" exactly fills the height of the new frame, either cropping
///       or letterboxing the width if it isn't precisely the right size to
///       preserve the original aspect ratio.
///     - "letterbox" (the default) chooses whichever of "width" or "height"
///       will maximally fill the new frame with no image data lost (it will
///       only letterbox, never crop).
///
/// If `exact` is true, will result in an exact match on aspect ratio and
/// centering (partial pixel shift if necessary), whereas exact=false
/// will only preserve aspect ratio and centering to the precision of a
/// whole pixel.
///
/// The filter is used to weight the `src` pixels falling underneath it for
/// each `dst` pixel.  The caller may specify a reconstruction filter by
/// name and width (expressed in pixels units of the `dst` image), or
/// `rotate()` will choose a reasonable default high-quality default filter
/// (lanczos3) if the empty string is passed, and a reasonable filter width
/// if `filterwidth` is 0. (Note that some filter choices only make sense
/// with particular width, in which case this filterwidth parameter may be
/// ignored.)
///
ImageBuf OIIO_API fit (const ImageBuf &src,
                       string_view filtername = "", float filterwidth=0.0f,
                       string_view fillmode="letterbox", bool exact=false,
                       ROI roi={}, int nthreads=0);
ImageBuf OIIO_API fit (const ImageBuf &src, Filter2D *filter,
                       string_view fillmode="letterbox", bool exact=false,
                       ROI roi={}, int nthreads=0);
bool OIIO_API fit (ImageBuf &dst, const ImageBuf &src,
                   string_view filtername = "", float filterwidth=0.0f,
                   string_view fillmode="letterbox", bool exact=false,
                   ROI roi={}, int nthreads=0);
bool OIIO_API fit (ImageBuf &dst, const ImageBuf &src, Filter2D *filter,
                   string_view fillmode="letterbox", bool exact=false,
                   ROI roi={}, int nthreads=0);

#ifndef DOXYGEN_SHOULD_SKIP_THIS
// DEPRECATED(2.3): old versions lacking the "fillmode" parameter
ImageBuf OIIO_API fit (const ImageBuf &src,
                       string_view filtername, float filterwidth,
                       bool exact, ROI roi={}, int nthreads=0);
ImageBuf OIIO_API fit (const ImageBuf &src, Filter2D *filter,
                       bool exact, ROI roi={}, int nthreads=0);
bool OIIO_API fit (ImageBuf &dst, const ImageBuf &src,
                   string_view filtername, float filterwidth,
                   bool exact, ROI roi={}, int nthreads=0);
bool OIIO_API fit (ImageBuf &dst, const ImageBuf &src, Filter2D *filter,
                   bool exact, ROI roi={}, int nthreads=0);
#endif
/// @}


/// @defgroup warp (warp: arbitrary warp by a 3x3 matrix)
/// @{
///
/// Warp the `src` image using the supplied 3x3 transformation matrix.
///
/// Only the pixels (and channels) of `dst` that are specified by `roi` will
/// be copied from the warped `src`; the default roi is to alter all the
/// pixels in dst. If `dst` is uninitialized, it will be sized to be an
/// ImageBuf large enough to hold the warped image if recompute_roi is true,
/// or will have the same ROI as src if recompute_roi is false. It is an
/// error to pass both an uninitialized `dst` and an undefined `roi`.
///
/// The caller may explicitly pass a reconstruction filter, or specify one
/// by name and size, or if the name is the empty string `resize()` will
/// choose a reasonable high-quality default if `nullptr` is passed.  The
/// filter is used to weight the `src` pixels falling underneath it for each
/// `dst` pixel; the filter's size is expressed in pixel units of the `dst`
/// image.

ImageBuf OIIO_API warp (const ImageBuf &src, M33fParam M,
                        string_view filtername = string_view(),
                        float filterwidth = 0.0f, bool recompute_roi = false,
                        ImageBuf::WrapMode wrap = ImageBuf::WrapDefault,
                        ROI roi={}, int nthreads=0);
ImageBuf OIIO_API warp (const ImageBuf &src, M33fParam M,
                        const Filter2D *filter, bool recompute_roi = false,
                        ImageBuf::WrapMode wrap = ImageBuf::WrapDefault,
                        ROI roi = {}, int nthreads=0);
bool OIIO_API warp (ImageBuf &dst, const ImageBuf &src, M33fParam M,
                    string_view filtername = string_view(),
                    float filterwidth = 0.0f, bool recompute_roi = false,
                    ImageBuf::WrapMode wrap = ImageBuf::WrapDefault,
                    ROI roi={}, int nthreads=0);
bool OIIO_API warp (ImageBuf &dst, const ImageBuf &src, M33fParam M,
                    const Filter2D *filter, bool recompute_roi = false,
                    ImageBuf::WrapMode wrap = ImageBuf::WrapDefault,
                    ROI roi = {}, int nthreads=0);

#ifdef OIIO_INTERNAL  /* experimental -- not part of public API yet */
ImageBuf OIIO_API warp (const ImageBuf &src, M33fParam M,
                        string_view filtername,
                        float filterwidth, bool recompute_roi,
                        ImageBuf::WrapMode wrap, bool edgeclamp,
                        ROI roi={}, int nthreads=0);
ImageBuf OIIO_API warp (const ImageBuf &src, M33fParam M,
                        const Filter2D *filter, bool recompute_roi,
                        ImageBuf::WrapMode wrap, bool edgeclamp,
                        ROI roi = {}, int nthreads=0);
bool OIIO_API warp (ImageBuf &dst, const ImageBuf &src, M33fParam M,
                    string_view filtername,
                    float filterwidth, bool recompute_roi,
                    ImageBuf::WrapMode wrap, bool edgeclamp,
                    ROI roi={}, int nthreads=0);
bool OIIO_API warp (ImageBuf &dst, const ImageBuf &src, M33fParam M,
                    const Filter2D *filter, bool recompute_roi,
                    ImageBuf::WrapMode wrap, bool edgeclamp,
                    ROI roi = {}, int nthreads=0);
#endif  // OIIO_INTERNAL
/// @}


/// @defgroup st_warp (st_warp: warp an image using per-pixel st coordinates)
/// @{
///
/// Warp the `src` image using "st" coordinates from a secondary `stbuf` image.
///
/// Each pixel in the `stbuf` image is used as a normalized image-space
/// coordinate in the `src` image, which is then sampled at that position using
/// the given reconstruction filter to produce an output pixel.
///
/// The transform is only defined over the area of the `stbuf` image, and thus
/// the given `roi` argument will be intersected with its geometry.
///
/// \b NOTE: The current behavior of this transform is modeled to match Nuke's
/// STMap node.
///
/// @param dst
///             The output ImageBuf. If an initialized buffer is provided, its
///             full-size dimensions must match those of `stbuf`.
/// @param src
///             The source ImageBuf to warp.
/// @param stbuf
///             The ImageBuf holding the st coordinates. This must be holding
///             a floating-point pixel data type.
/// @param chan_s
///             The index of the "s" channel in the `stbuf` image. This defaults
///             to its first channel.
/// @param chan_t
///             The index of the "t" channel in the `stbuf` image. This defaults
///             to its second channel.
/// @param flip_s
///             Whether to mirror the "s" coordinate along the horizontal axis
///             when computing source pixel positions. This is useful if the
///             coordinates are defined in terms of a different image origin 
///             than OpenImageIO's.
/// @param flip_t
///             Whether to mirror the "t" coordinate along the vertical axis
///             when computing source pixel positions. This is useful if the
///             coordinates are defined in terms of a different image origin 
///             than OpenImageIO's.

ImageBuf OIIO_API st_warp (const ImageBuf &src, const ImageBuf& stbuf,
                           string_view filtername=string_view(),
                           float filterwidth=0.0f, int chan_s=0, int chan_t=1,
                           bool flip_s=false, bool flip_t=false, ROI roi={},
                           int nthreads=0);
ImageBuf OIIO_API st_warp (const ImageBuf &src, const ImageBuf& stbuf,
                           const Filter2D *filter, int chan_s=0, int chan_t=1,
                           bool flip_s=false, bool flip_t=false, ROI roi={},
                           int nthreads=0);
bool OIIO_API st_warp (ImageBuf &dst, const ImageBuf &src,
                       const ImageBuf& stbuf,
                       string_view filtername=string_view(),
                       float filterwidth=0.0f, int chan_s=0, int chan_t=1,
                       bool flip_s=false, bool flip_t=false, ROI roi={},
                       int nthreads=0);
bool OIIO_API st_warp (ImageBuf &dst, const ImageBuf &src,
                       const ImageBuf& stbuf, const Filter2D *filter,
                       int chan_s=0, int chan_t=1, bool flip_s=false,
                       bool flip_t=false, ROI roi={}, int nthreads=0);
/// @}


/// Compute per-pixel sum `A + B`, returning the result image.
///
/// `A` and `B` may each either be an `ImageBuf&`, or a `cspan<float>`
/// giving a per- channel constant, or a single constant used for all
/// channels. (But at least one must be an image.)
ImageBuf OIIO_API add (Image_or_Const A, Image_or_Const B,
                       ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API add (ImageBuf &dst, Image_or_Const A, Image_or_Const B,
                   ROI roi={}, int nthreads=0);


/// Compute per-pixel signed difference `A - B`, returning the result image.
///
/// `A` and `B` may each either be an `ImageBuf&`, or a `cspan<float>`
/// giving a per-channel constant, or a single constant used for all
/// channels. (But at least one must be an image.)
ImageBuf OIIO_API sub (Image_or_Const A, Image_or_Const B,
                       ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API sub (ImageBuf &dst, Image_or_Const A, Image_or_Const B,
                   ROI roi={}, int nthreads=0);


/// Compute per-pixel absolute difference `abs(A - B)`, returning the result
/// image.
///
/// `A` and `B` may each either be an `ImageBuf&`, or a `cspan<float>`
/// giving a per- channel constant, or a single constant used for all
/// channels. (But at least one must be an image.)
ImageBuf OIIO_API absdiff (Image_or_Const A, Image_or_Const B,
                           ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API absdiff (ImageBuf &dst, Image_or_Const A, Image_or_Const B,
                       ROI roi={}, int nthreads=0);


/// Compute per-pixel absolute value `abs(A)`, returning the result image.
ImageBuf OIIO_API abs (const ImageBuf &A, ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API abs (ImageBuf &dst, const ImageBuf &A, ROI roi={}, int nthreads=0);


/// Compute per-pixel product `A * B`, returning the result image.
///
/// Either both `A` and `B` are images, or one is an image and the other is
/// a `cspan<float>` giving a per-channel constant or a single constant
/// used for all channels.
ImageBuf OIIO_API mul (Image_or_Const A, Image_or_Const B,
                       ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API mul (ImageBuf &dst, Image_or_Const A, Image_or_Const B,
                   ROI roi={}, int nthreads=0);


/// Compute per-pixel division `A / B`, returning the result image.
/// Division by zero is defined to result in zero.
///
/// `A` is always an image, and `B` is either an image or a `cspan<float>`
/// giving a per-channel constant or a single constant used for all
/// channels.
ImageBuf OIIO_API div (Image_or_Const A, Image_or_Const B,
                       ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API div (ImageBuf &dst, Image_or_Const A, Image_or_Const B,
                   ROI roi={}, int nthreads=0);


/// Compute per-pixel multiply-and-add `A * B + C`, returning the result
/// image.
///
/// `A`, `B`, and `C` are each either an image, or a `cspan<float>` giving a
/// per-channel constant or a single constant used for all channels. (Note:
/// at least one must be an image.)
ImageBuf OIIO_API mad (Image_or_Const A, Image_or_Const B,
                       Image_or_Const C, ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API mad (ImageBuf &dst, Image_or_Const A, Image_or_Const B,
                   Image_or_Const C, ROI roi={}, int nthreads=0);


/// Return the composite of `A` over `B` using the Porter/Duff definition of
/// "over", returning true upon success and false for any of a variety of
/// failures (as described below).
///
/// `A` and `B` (and dst, if already defined/allocated) must have valid
/// alpha channels identified by their ImageSpec `alpha_channel` field.  If`
/// A` or `B` do not have alpha channels (as determined by those rules) or
/// if the number of non-alpha channels do not match between `A` and `B`,
/// `over()` will fail, returning false.
///
/// If `dst` is not already an initialized ImageBuf, it will be sized to
/// encompass the minimal rectangular pixel region containing the union of
/// the defined pixels of `A` and `B`, and with a number of channels equal
/// to the number of non-alpha channels of `A` and `B`, plus an alpha
/// channel.  However, if `dst` is already initialized, it will not be
/// resized, and the "over" operation will apply to its existing pixel data
/// window.  In this case, dst must have an alpha channel designated and
/// must have the same number of non-alpha channels as `A` and `B`,
/// otherwise it will fail, returning false.
///
/// `A`, `B`, and `dst` need not perfectly overlap in their pixel data
/// windows; pixel values of `A` or `B` that are outside their respective
/// pixel data window will be treated as having "zero" (0,0,0...) value.
ImageBuf OIIO_API over (const ImageBuf &A, const ImageBuf &B,
                        ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API over (ImageBuf &dst, const ImageBuf &A, const ImageBuf &B,
                    ROI roi={}, int nthreads=0);


/// Just like `ImageBufAlgo::over()`, but inputs `A` and `B` must have
/// designated 'z' channels, and on a pixel-by-pixel basis, the z values
/// will determine which of `A` or `B` will be considered the foreground or
/// background (lower z is foreground).  If `z_zeroisinf` is true, then z=0
/// values will be treated as if they are infinitely far away.
ImageBuf OIIO_API zover (const ImageBuf &A, const ImageBuf &B,
                         bool z_zeroisinf=false, ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API zover (ImageBuf &dst, const ImageBuf &A, const ImageBuf &B,
                     bool z_zeroisinf=false, ROI roi={}, int nthreads=0);



/// Compute per-pixel value inverse `1.0 - A` (which you can think of as
/// roughly meaning switching white and black), returning the result image.
///
/// Tips for callers: (1) You probably want to set `roi` to restrict the
/// operation to only the color channels, and not accidentally include
/// alpha, z, or others. (2) There may be situations where you want to
/// `unpremult()` before the invert, then `premult()` the result, so that
/// you are computing the inverse of the unmasked color.
ImageBuf OIIO_API invert (const ImageBuf &A, ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API invert (ImageBuf &dst, const ImageBuf &A, ROI roi={}, int nthreads=0);


/// Compute per-pixel raise-to-power `A ^ B`. returning the result image. It
/// is permitted for `dst` and `A` to be the same image.
///
/// `A` is always an image, and `B` is either an image or a `cspan<float>`
/// giving a per-channel constant or a single constant used for all
/// channels.
ImageBuf OIIO_API pow (const ImageBuf &A, cspan<float> B,
                       ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API pow (ImageBuf &dst, const ImageBuf &A, cspan<float> B,
                   ROI roi={}, int nthreads=0);

/// Normalize a 3D vector texture (i.e., divide each pixel by its length).
/// This function assumes a 3-channel image that represents a 3-vector, or a
/// 4-channel image that represents a 3-vector plus an alpha value. If an
/// alpha channel is present, its value is merely copied, and is not part of
/// the normalization computation. If the destination has no alpha channel but
/// the sources do, the alpha channel will be dropped.
///
/// `inCenter` and `outCenter` define the pixel value that corresponds to a
/// 0.0 vector value for input and output, respectively.  `scale` defines the
/// scale factor to apply to the normalized vectors.
///
/// Thus, if the input image encodes vector components into [0,1] range pixel
/// values so that a pixel value 0.5 indicates a 0-length vector, then you
/// should use `inCenter=0.5`, whereas if they are already using the full
/// range (0.0 is encoded as 0.0), then you want `inCenter=0.0`. Similarly, if
/// you want the output normalized vectors to be in the range [0,1], use
/// `outCenter=0.5` and `scale=0.5`, but if you want them to be in the range
/// [-1,1], use `outCenter=0.0` and `scale=1.0` (this probably will only work
/// if you intend to write the results in `float` or `half` format).
///
/// Expressed another way, the computation is conceptually:
///
///     out = outCenter + scale * (in - inCenter) / length(in - inCenter)
/// 
bool OIIO_API normalize(ImageBuf& dst, const ImageBuf& A, float inCenter=0.0f,
                        float outCenter=0.0f, float scale=1.0f,
                        ROI roi={}, int nthreads=0);

ImageBuf OIIO_API normalize(const ImageBuf& A, float inCenter=0.0f,
                            float outCenter=0.0, float scale=1.0f, 
                            ROI roi={}, int nthreads=0);


/// Converts a multi-channel image into a one-channel image via a weighted
/// sum of channels:
///
///     (channel[0]*weight[0] + channel[1]*weight[1] + ...)
///
/// returning the resulting one-channel image. The `weights`, if not
/// supplied, default to `{ 1, 1, 1, ... }`).
ImageBuf OIIO_API channel_sum (const ImageBuf &src, cspan<float> weights=1.0f,
                               ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API channel_sum (ImageBuf &dst, const ImageBuf &src,
                           cspan<float> weights=1.0f,
                           ROI roi={}, int nthreads=0);


/// @defgroup maxminclamp (Maximum, minimum, clamping)
/// @{
///
/// `max()` and `min()` take the pixel-by-pixel, channel-by-channel
/// maximum and minimum of two images, or of an image and a constant.
///
/// `clamp()` restricts values of an image to the range between per-channel
/// minimum and maximum constant values.


/// Compute per-pixel `max(A, B)`, returning the result image.
///
/// Either both `A` and `B` are images, or one is an image and the other is
/// a `cspan<float>` giving a per-channel constant or a single constant
/// used for all channels.
ImageBuf OIIO_API max (Image_or_Const A, Image_or_Const B,
                       ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API max (ImageBuf &dst, Image_or_Const A, Image_or_Const B,
                   ROI roi={}, int nthreads=0);

/// Compute per-pixel `min(A, B)`, returning the result image.
///
/// Either both `A` and `B` are images, or one is an image and the other is
/// a `cspan<float>` giving a per-channel constant or a single constant
/// used for all channels.
ImageBuf OIIO_API min (Image_or_Const A, Image_or_Const B,
                       ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API min (ImageBuf &dst, Image_or_Const A, Image_or_Const B,
                   ROI roi={}, int nthreads=0);



/// Return pixels of `src` with pixel values clamped as follows:
/// @param min
///              The minimum clamp value for each channel. If `min` is
///              empty, no minimum clamping is performed.
/// @param max
///              The maximum clamp value for each channel. If `max` is
///              empty, no maximum clamping is performed.
/// @param clampalpha01
///              If true, then additionally any alpha channel is clamped
///              to the 0-1 range.
ImageBuf OIIO_API clamp (const ImageBuf &src,
                         cspan<float> min=-std::numeric_limits<float>::max(),
                         cspan<float> max=std::numeric_limits<float>::max(),
                         bool clampalpha01 = false, ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API clamp (ImageBuf &dst, const ImageBuf &src,
                     cspan<float> min=-std::numeric_limits<float>::max(),
                     cspan<float> max=std::numeric_limits<float>::max(),
                     bool clampalpha01 = false, ROI roi={}, int nthreads=0);

/// @}


/// @defgroup maxminchan (Maximum / minimum of channels)
/// @{
///
/// `maxchan()` computes a one-channel image that for each pixel, contains the
/// maximum value of all channels of corresponding pixel of the source image.
/// `minchan()` similarly computes the minimum value of all channels.
///
/// @version 2.3.10

ImageBuf OIIO_API maxchan (const ImageBuf& A, ROI roi={}, int nthreads=0);
bool OIIO_API maxchan (ImageBuf &dst, const ImageBuf& A,
                       ROI roi={}, int nthreads=0);

ImageBuf OIIO_API minchan (const ImageBuf& src, ROI roi={}, int nthreads=0);
bool OIIO_API minchan (ImageBuf &dst, const ImageBuf& src,
                       ROI roi={}, int nthreads=0);
/// @}


/// Return pixel values that are a contrast-remap of the corresponding
/// values of the `src` image, transforming pixel value domain [black,
/// white] to range [min, max], either linearly or with optional application
/// of a smooth sigmoidal remapping (if `scontrast` != 1.0).
///
/// The following steps are performed, in order:
///
/// 1. Linearly rescale values [`black`, `white`] to [0, 1].
/// 2. If `scontrast` != 1, apply a sigmoidal remapping where a larger
///    `scontrast` value makes a steeper slope, and the steepest part is at
///    value `sthresh` (relative to the new remapped value after steps 1 &
///    2; the default is 0.5).
/// 3. Rescale the range of that result: 0.0 -> `min` and 1.0 -> `max`.
///
/// Values outside of the [black,white] range will be extrapolated to
/// outside [min,max], so it may be prudent to apply a clamp() to the
/// results.
///
/// The black, white, min, max, scontrast, sthresh parameters may each
/// either be a single float value for all channels, or a span giving
/// per-channel values.
///
/// You can use this function for a simple linear contrast remapping of
/// [black, white] to [min, max] if you use the default values for sthresh.
/// Or just a simple sigmoidal contrast stretch within the [0,1] range if
/// you leave all other parameters at their defaults, or a combination of
/// these effects. Note that if `black` == `white`, the result will be a
/// simple binary thresholding where values < `black` map to `min` and
/// values >= `black` map to `max`.

OIIO_API ImageBuf contrast_remap (const ImageBuf &src,
                    cspan<float> black=0.0f, cspan<float> white=1.0f,
                    cspan<float> min=0.0f, cspan<float> max=1.0f,
                    cspan<float> scontrast=1.0f, cspan<float> sthresh=0.5f,
                    ROI={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
OIIO_API bool contrast_remap (ImageBuf &dst, const ImageBuf &src,
                    cspan<float> black=0.0f, cspan<float> white=1.0f,
                    cspan<float> min=0.0f, cspan<float> max=1.0f,
                    cspan<float> scontrast=1.0f, cspan<float> sthresh=0.5f,
                    ROI={}, int nthreads=0);


/// @defgroup saturate (Adjust saturation of color channels)
/// @{
///
/// Increase or decrease color saturation of the image.
///
/// The `saturate` operation returns (or copies into `dst`) the pixels of
/// `src` within the ROI, and in the process adjusts the color saturation of
/// the three consecutive channels starting with `firstchannel` based on the
/// `scale` parameter: 0.0 fully desaturates to a greyscale image of
/// percaptually equivalent luminance, 1.0 leaves the colors unchanged,
/// `scale` values inside this range interpolate between them, and `scale` > 1
/// would increase apparent color saturation.
///
/// Channels that are within the range of `roi.chbegin` to `roi.chend-1`, but
/// outside the range of `firstchannel` to `firstchannel+2` are simply copied
/// unaltered. Only three channels at a time can be desaturated, by default
/// the first three channels, though `firstchannel` may be used to specify a
/// different subset of channels. It is allowed for `src` and `dst` to be the
/// same image.
///
/// @version 2.4+

ImageBuf OIIO_API saturate (const ImageBuf &src, float scale = 0.0f,
                            int firstchannel = 0, ROI roi={}, int nthreads=0);
bool OIIO_API saturate (ImageBuf &dst, const ImageBuf &src, float scale = 0.0f,
                        int firstchannel = 0, ROI roi={}, int nthreads=0);
/// @}



/// @defgroup color_map (Remap value range by spline or name)
/// @{
///
/// Remap value range by spline or name
///
/// Return (or copy into `dst`) pixel values determined by looking up a
/// color map using values of the source image, using either the channel
/// specified by `srcchannel`, or the luminance of `src`'s RGB if
/// `srcchannel` is -1. This happens for all pixels within the  ROI (which
/// defaults to all of `src`), and if `dst` is not already initialized, it
/// will be initialized to the ROI and with color channels equal to
/// `channels`.
///
/// In the variant that takes a `knots` parameter, this specifies the values
/// of a linearly-interpolated color map given by `knots[nknots*channels]`.
/// An input value of 0.0 is mapped to `knots[0..channels-1]` (one value for
/// each color channel), and an input value of 1.0 is mapped to
/// `knots[(nknots-1)*channels..knots.size()-1]`.
///
/// In the variant that takes a `mapname` parameter, this is the name of a
/// color map. Recognized map names include: "inferno", "viridis", "magma",
/// "plasma", all of which are perceptually uniform, strictly increasing in
/// luminance, look good when converted to grayscale, and work for people
/// with all types of colorblindness. Also "turbo" has most of these
/// properties (except for being strictly increasing in luminance) and
/// is a nice rainbow-like pattern. Also supported are the following color
/// maps that do not have those desirable qualities (and are thus not
/// recommended, but are present for back-compatibility or for use by
/// clueless people): "blue-red", "spectrum", and "heat". In all cases, the
/// implied `channels` is 3.

ImageBuf OIIO_API color_map (const ImageBuf &src, int srcchannel,
                             int nknots, int channels, cspan<float> knots,
                             ROI roi={}, int nthreads=0);
ImageBuf OIIO_API color_map (const ImageBuf &src, int srcchannel,
                             string_view mapname, ROI roi={}, int nthreads=0);
bool OIIO_API color_map (ImageBuf &dst, const ImageBuf &src, int srcchannel,
                         int nknots, int channels, cspan<float> knots,
                         ROI roi={}, int nthreads=0);
bool OIIO_API color_map (ImageBuf &dst, const ImageBuf &src, int srcchannel,
                         string_view mapname, ROI roi={}, int nthreads=0);
/// @}


/// @defgroup range (Nonlinear range remapping for contrast preservation)
/// @{
///
/// Nonlinear range remapping for contrast preservation
///
/// `rangecompress()` returns (or copy into `dst`) all pixels and color
/// channels of `src` within region `roi` (defaulting to all the defined
/// pixels of `dst`), rescaling their range with a logarithmic
/// transformation. Alpha and z channels are not transformed.
///
/// `rangeexpand()` performs the inverse transformation (logarithmic back
/// into linear).
///
/// If `useluma` is true, the luma of channels [roi.chbegin..roi.chbegin+2]
/// (presumed to be R, G, and B) are used to compute a single scale factor
/// for all color channels, rather than scaling all channels individually
/// (which could result in a color shift).
///
/// The purpose of these function is as follows: Some image operations (such
/// as resizing with a "good" filter that contains negative lobes) can have
/// objectionable artifacts when applied to images with very high-contrast
/// regions involving extra bright pixels (such as highlights in HDR
/// captured or rendered images).  By compressing the range pixel values,
/// then performing the operation, then expanding the range of the result
/// again, the result can be much more pleasing (even if not exactly
/// correct).

ImageBuf OIIO_API rangecompress (const ImageBuf &src, bool useluma = false,
                                 ROI roi={}, int nthreads=0);
ImageBuf OIIO_API rangeexpand (const ImageBuf &src, bool useluma = false,
                               ROI roi={}, int nthreads=0);
bool OIIO_API rangecompress (ImageBuf &dst, const ImageBuf &src,
                             bool useluma = false, ROI roi={}, int nthreads=0);
bool OIIO_API rangeexpand (ImageBuf &dst, const ImageBuf &src,
                           bool useluma = false, ROI roi={}, int nthreads=0);
/// @}


struct OIIO_API PixelStats {
    std::vector<float> min;
    std::vector<float> max;
    std::vector<float> avg;
    std::vector<float> stddev;
    std::vector<imagesize_t> nancount;
    std::vector<imagesize_t> infcount;
    std::vector<imagesize_t> finitecount;
    std::vector<double> sum, sum2;  // for intermediate calculation

    PixelStats () {}
    PixelStats (PixelStats&& other) = default;
    PixelStats (int nchannels) { reset(nchannels); }
    void reset (int nchannels);
    void merge (const PixelStats &p);
    const PixelStats& operator= (PixelStats&& other);  // Move assignment
};


/// Compute statistics about the ROI of the `src` image, returning a
/// PixelStats structure. Upon success, the returned vectors in the result
/// structure will have size == src.nchannels(). If there is a failure, the
/// vector sizes will be 0 and an error will be set in src.
PixelStats OIIO_API computePixelStats (const ImageBuf &src,
                                       ROI roi={}, int nthreads=0);

// DEPRECATED(1.9): with C++11 move semantics, there's no reason why
// stats needs to be passed as a parameter instead of returned.
#ifndef DOXYGEN_SHOULD_SKIP_THIS
OIIO_DEPRECATED("use version that returns PixelStats (1.9)")
bool OIIO_API computePixelStats (PixelStats &stats, const ImageBuf &src,
                                 ROI roi={}, int nthreads=0);
#endif


// Struct holding all the results computed by ImageBufAlgo::compare().
// (maxx,maxy,maxz,maxc) gives the pixel coordinates (x,y,z) and color
// channel of the pixel that differed maximally between the two images.
// nwarn and nfail are the number of "warnings" and "failures",
// respectively.
struct CompareResults {
    double meanerror, rms_error, PSNR, maxerror;
    int maxx, maxy, maxz, maxc;
    imagesize_t nwarn, nfail;
    bool error;
};

/// Numerically compare two images.  The difference threshold (for any
/// individual color channel in any pixel) for a "failure" is `failthresh`,
/// and for a "warning" is `warnthresh`.  If nonzero, then `failrelative` and
/// `warnrelative` are alternate thresholds as a portion the mean of the
/// absolute values of the two images. It only warns or fails if both criteria
/// are met. More formally, a value comparison will fail if
///
///     abs(A-B) > failthresh && abs(A-B)/((abs(A)+abs(B))/2) > failrelative
///
/// and analogously for warning.
///
/// The results are stored in `result`.  If `roi` is defined, pixels will be
/// compared for the pixel and channel range that is specified.  If `roi` is
/// not defined, the comparison will be for all channels, on the union of the
/// defined pixel windows of the two images (for either image, undefined
/// pixels will be assumed to be black).
CompareResults OIIO_API compare (const ImageBuf &A, const ImageBuf &B,
                                 float failthresh, float warnthresh,
                                 float failrelative, float warnrelative,
                                 ROI roi={}, int nthreads=0);

/// Numerically compare two images.  The difference threshold (for any
/// individual color channel in any pixel) for a "failure" is
/// failthresh, and for a "warning" is warnthresh.  The results are
/// stored in result.  If roi is defined, pixels will be compared for
/// the pixel and channel range that is specified.  If roi is not
/// defined, the comparison will be for all channels, on the union of
/// the defined pixel windows of the two images (for either image,
/// undefined pixels will be assumed to be black).
CompareResults OIIO_API compare (const ImageBuf &A, const ImageBuf &B,
                                 float failthresh, float warnthresh,
                                 ROI roi={}, int nthreads=0);

/// Compare two images using Hector Yee's perceptual metric, returning
/// the number of pixels that fail the comparison.  Only the first three
/// channels (or first three channels specified by `roi`) are compared.
/// Free parameters are the ambient luminance in the room and the field
/// of view of the image display; our defaults are probably reasonable
/// guesses for an office environment.  The 'result' structure will
/// store the maxerror, and the maxx, maxy, maxz of the pixel that
/// failed most severely.  (The other fields of the CompareResults
/// are not used for Yee comparison.)
///
/// Works for all pixel types.  But it's basically meaningless if the
/// first three channels aren't RGB in a linear color space that sort
/// of resembles AdobeRGB.
///
/// Return true on success, false on error.
int OIIO_API compare_Yee (const ImageBuf &A, const ImageBuf &B,
                          CompareResults &result,
                          float luminance = 100, float fov = 45,
                          ROI roi={}, int nthreads=0);

#ifndef DOXYGEN_SHOULD_SKIP_THIS
// DEPRECATED(1.9): with C++11 move semantics, there's no reason why
// result needs to be passed as a parameter instead of returned.
OIIO_DEPRECATED("use version that returns CompareResults (1.9)")
bool OIIO_API compare (const ImageBuf &A, const ImageBuf &B,
                       float failthresh, float warnthresh,
                       CompareResults &result, ROI roi={}, int nthreads=0);
#endif


/// Do all pixels within the ROI have the same values for channels
/// `[roi.chbegin..roi.chend-1]`, within a tolerance of +/- `threshold`?  If
/// so, return `true` and store that color in `color[chbegin...chend-1]` (if
/// `color` is not empty); otherwise return `false`.  If `roi` is not
/// defined (the default), it will be understood to be all of the defined
/// pixels and channels of source.
OIIO_API bool isConstantColor (const ImageBuf &src, float threshold=0.0f,
                               span<float> color = {},
                               ROI roi={}, int nthreads=0);
#ifndef DOXYGEN_SHOULD_SKIP_THIS
inline bool isConstantColor (const ImageBuf &src, span<float> color,
                             ROI roi={}, int nthreads=0) {
    return isConstantColor (src, 0.0f, color, roi, nthreads);
}
#endif


/// Does the requested channel have a given value (within a tolerance of +/-
/// `threshold`) for every channel within the ROI?  (For this function, the
/// ROI's chbegin/chend are ignored.)  Return `true` if so, otherwise return
/// `false`.  If `roi` is not defined (the default), it will be understood
/// to be all of the defined pixels and channels of source.
OIIO_API bool isConstantChannel (const ImageBuf &src, int channel,
                                 float val, float threshold=0.0f,
                                 ROI roi={}, int nthreads=0);
#ifndef DOXYGEN_SHOULD_SKIP_THIS
inline bool isConstantChannel (const ImageBuf &src, int channel, float val,
                               ROI roi, int nthreads=0) {
    return isConstantChannel (src, channel, val, 0.0f, roi, nthreads);
}
#endif


/// Is the image monochrome within the ROI, i.e., for every pixel within the
/// region, do all channels [roi.chbegin, roi.chend) have the same value
/// (within a tolerance of +/- threshold)?  If roi is not defined (the
/// default), it will be understood to be all of the defined pixels and
/// channels of source.
OIIO_API bool isMonochrome (const ImageBuf &src, float threshold=0.0f,
                            ROI roi={}, int nthreads=0);
#ifndef DOXYGEN_SHOULD_SKIP_THIS
inline bool isMonochrome (const ImageBuf &src, ROI roi, int nthreads=0) {
    return isMonochrome (src, 0.0f, roi, nthreads);
}
#endif


/// Count how many pixels in the ROI match a list of colors. The colors to
/// match are in:
/// 
///     colors[0 ... nchans-1]
///     colors[nchans ... 2*nchans-1]
///     ...
///     colors[(ncolors-1)*nchans ... (ncolors*nchans)-1]
/// 
/// and so on, a total of `ncolors` consecutively stored colors of `nchans`
/// channels each (`nchans` is the number of channels in the image, itself,
/// it is not passed as a parameter).
///
/// `eps[0..nchans-1]` are the error tolerances for a match, for each
/// channel. Setting `eps[c]` = `numeric_limits<float>::max()` will
/// effectively make it ignore the channel.  The default `eps` is 0.001 for
/// all channels (this value is chosen because it requires exact matches for
/// 8 bit images, but allows a wee bit of imprecision for float images.
///
/// Upon success, return `true` and store the number of pixels that matched
/// each color `count[0..ncolors-1]`.  If there is an error, returns `false`
/// and sets an appropriate error message set in `src`.
bool OIIO_API color_count (const ImageBuf &src, imagesize_t *count,
                           int ncolors, cspan<float> color,
                           cspan<float> eps = 0.001f,
                           ROI roi={}, int nthreads=0);


/// Count how many pixels in the image (within the ROI) are outside the
/// value range described by `low[roi.chbegin..roi.chend-1]` and
/// `high[roi.chbegin..roi.chend-1]` as the low and high acceptable values
/// for each color channel.
/// 
/// The number of pixels containing values that fall below the lower bound
/// will be stored in `*lowcount`, the number of pixels containing
/// values that fall above the upper bound will be stored in 
/// `*highcount`, and the number of pixels for which all channels fell
/// within the bounds will be stored in `*inrangecount`.  Any of these
/// may be NULL, which simply means that the counts need not be collected or
/// stored.
bool OIIO_API color_range_check (const ImageBuf &src,
                                 imagesize_t *lowcount,
                                 imagesize_t *highcount,
                                 imagesize_t *inrangecount,
                                 cspan<float> low, cspan<float> high,
                                 ROI roi={}, int nthreads=0);

/// Find the minimal rectangular region within `roi` (which defaults to the
/// entire pixel data window of `src`) that consists of nonzero pixel
/// values.  In other words, gives the region that is a "shrink-wraps" of
/// `src` to exclude black border pixels.  Note that if the entire image was
/// black, the ROI returned will contain no pixels.
///
/// For "deep" images, this function returns the smallest ROI that contains
/// all pixels that contain depth samples, and excludes the border pixels
/// that contain no depth samples at all.
OIIO_API ROI nonzero_region (const ImageBuf &src, ROI roi={}, int nthreads=0);


/// Compute the SHA-1 byte hash for all the pixels in the specified region of
/// the image.  If `blocksize` > 0, the function will compute separate SHA-1
/// hashes of each `blocksize` batch of scanlines, then return a hash of the
/// individual hashes.  This is just as strong a hash, but will NOT match a
/// single hash of the entire image (`blocksize==0`).  But by breaking up
/// the hash into independent blocks, we can parallelize across multiple
/// threads, given by `nthreads` (if `nthreads` is 0, it will use the global
/// OIIO thread count).  The `extrainfo` provides additional text that will
/// be incorporated into the hash.
std::string OIIO_API computePixelHashSHA1 (const ImageBuf &src,
                                           string_view extrainfo = "",
                                           ROI roi={},
                                           int blocksize = 0, int nthreads=0);


/// Compute a histogram of `src`, for the given channel and ROI. Return a
/// vector of length `bins` that contains the counts of how many pixel
/// values were in each of `bins` equally spaced bins covering the range of
/// values `[min,max]`. Values < `min` count for bin 0, values > `max` count
/// for bin `nbins-1`. If `ignore_empty` is `true`, no counts will be
/// incremented for any pixels whose value is 0 in all channels.
///
/// If there was an error, the returned vector will be empty, and an error
/// message will be retrievable from src.geterror().
OIIO_API
std::vector<imagesize_t> histogram (const ImageBuf &src, int channel=0,
                                    int bins=256, float min=0.0f, float max=1.0f,
                                    bool ignore_empty=false,
                                    ROI roi={}, int nthreads=0);


#ifndef DOXYGEN_SHOULD_SKIP_THIS
/// DEPRECATED(1.9)
OIIO_DEPRECATED("use version that returns vector (1.9)")
bool OIIO_API histogram (const ImageBuf &src, int channel,
                         std::vector<imagesize_t> &histogram, int bins=256,
                         float min=0, float max=1, imagesize_t *submin=nullptr,
                         imagesize_t *supermax=nullptr, ROI roi={});

// DEPRECATED(1.9): never liked this.
OIIO_DEPRECATED("this useless function is going away (1.9)")
bool OIIO_API histogram_draw (ImageBuf &dst,
                              const std::vector<imagesize_t> &histogram);
#endif



/// Make a 1-channel `float` image of the named kernel. The size of the
/// image will be big enough to contain the kernel given its size (`width` x
/// `height`) and rounded up to odd resolution so that the center of the
/// kernel can be at the center of the middle pixel.  The kernel image will
/// be offset so that its center is at the (0,0) coordinate.  If `normalize`
/// is true, the values will be normalized so that they sum to 1.0. If
/// `depth` > 1, a volumetric kernel will be created.  Use with caution!
///
/// Kernel names can be: "gaussian", "sharp-gaussian", "box",
/// "triangle", "blackman-harris", "mitchell", "b-spline", "catmull-rom",
/// "lanczos3", "disk", "binomial", "laplacian".
///
/// Note that "catmull-rom" and "lanczos3" are fixed-size kernels that
/// don't scale with the width, and are therefore probably less useful
/// in most cases.
///
/// The ImageBuf that is returned indicates if there was an error, in which
/// case return.has_error() will be true and return.geterror() can be used
/// to retrieve an error message.
ImageBuf OIIO_API make_kernel (string_view name, float width, float height,
                               float depth = 1.0f, bool normalize = true);
#ifndef DOXYGEN_SHOULD_SKIP_THIS
// DEPRECATED(1.9):
OIIO_DEPRECATED("use version that returns ImageBuf (1.9)")
inline bool make_kernel (ImageBuf &dst, string_view name,
                         float width, float height, float depth = 1.0f,
                         bool normalize = true) {
    dst = make_kernel (name, width, height, depth, normalize);
    return ! dst.has_error();
}
#endif


/// Return the convolution of `src` and a `kernel`. If `roi` is not defined,
/// it defaults to the full size `src`. If `normalized` is true, the kernel will
/// be normalized for the  convolution, otherwise the original values will
/// be used.
ImageBuf OIIO_API convolve (const ImageBuf &src, const ImageBuf &kernel,
                            bool normalize = true, ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
/// If `roi` is not defined, it defaults to the full size of `dst` (or
/// `src`, if `dst` was uninitialized).  If `dst` is uninitialized, it will
/// be allocated to be the size specified by `roi`.
bool OIIO_API convolve (ImageBuf &dst, const ImageBuf &src, const ImageBuf &kernel,
                        bool normalize = true, ROI roi={}, int nthreads=0);


/// Return the Laplacian of the corresponding region of `src`.  The
/// Laplacian is the generalized second derivative of the image
/// \f[
/// \frac{\partial^2 s}{\partial x^2} + \frac{\partial^2 s}{\partial y^2}
/// \f]
/// which is approximated by convolving the image with a discrete 3x3
/// Laplacian kernel,
///
///                     [ 0  1  0 ]
///                     [ 1 -4  1 ]
///                     [ 0  1  0 ]
///
ImageBuf OIIO_API laplacian (const ImageBuf &src, ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API laplacian (ImageBuf &dst, const ImageBuf &src,
                         ROI roi={}, int nthreads=0);


/// @defgroup fft-ifft (Fast Fourier Transform and inverse)
/// @{
///
/// Fast Fourier Transform and inverse
///
/// Return (or copy into `dst`) the discrete Fourier transform (DFT), or its
/// inverse, of the section of `src` denoted by roi,  If roi is not defined,
/// it will be all of `src`'s pixels.
///
/// `fft()` takes the discrete Fourier transform (DFT) of the section of
/// `src` denoted by `roi`, returning it or storing it in `dst`. If `roi` is
/// not defined, it will be all of `src`'s pixels. Only one channel of `src`
/// may be transformed at a time, so it will be the first channel described
/// by `roi` (or, again, channel 0 if `roi` is undefined).  If not already
/// in the correct format, `dst` will be re-allocated to be a 2-channel
/// `float` buffer of size `roi.width()` x `roi.height`, with channel 0
/// being the "real" part and channel 1 being the the "imaginary" part.  The
/// values returned are actually the unitary DFT, meaning that it is scaled
/// by 1/sqrt(npixels).
///
/// `ifft()` takes the inverse discrete Fourier transform, transforming a
/// 2-channel complex (real and imaginary) frequency domain image and into a
/// single-channel spatial domain image. `src` must be a 2-channel float
/// image, and is assumed to be a complex frequency-domain signal with the
/// "real" component in channel 0 and the "imaginary" component in channel 1.
/// `dst` will end up being a float image of one channel (the real component
/// is kept, the imaginary component of the spatial-domain will be
/// discarded). Just as with `fft()`, the `ifft()` function is dealing with
/// the unitary DFT, so it is scaled by 1/sqrt(npixels).
ImageBuf OIIO_API fft (const ImageBuf &src, ROI roi={}, int nthreads=0);
ImageBuf OIIO_API ifft (const ImageBuf &src, ROI roi={}, int nthreads=0);
bool OIIO_API fft (ImageBuf &dst, const ImageBuf &src, ROI roi={}, int nthreads=0);
bool OIIO_API ifft (ImageBuf &dst, const ImageBuf &src, ROI roi={}, int nthreads=0);
/// @}


/// @defgroup complex-polar (Converting complex to polar and back)
/// @{
///
/// Converting complex to polar and back
///
/// The `polar_to_complex()` function transforms a 2-channel image whose
/// channels are interpreted as complex values (real and imaginary
/// components) into the equivalent values expressed in polar form of
/// amplitude and phase (with phase between 0 and \f$ 2\pi \f$.
/// 
/// The `complex_to_polar()` function performs the reverse transformation,
/// converting from  polar values (amplitude and phase) to complex (real and
/// imaginary).
/// 
/// In either case,  the section of `src` denoted by `roi` is transformed,
/// storing the result in `dst`. If `roi` is not defined, it will be all of
/// `src`'s pixels.  Only the first two channels of `src` will be
/// transformed.
/// 
/// The transformation between the two representations are:
///
///     real = amplitude * cos(phase);
///     imag = amplitude * sin(phase);
///
///     amplitude = hypot (real, imag);
///     phase = atan2 (imag, real);

ImageBuf OIIO_API complex_to_polar (const ImageBuf &src,
                                    ROI roi={}, int nthreads=0);
bool OIIO_API complex_to_polar (ImageBuf &dst, const ImageBuf &src,
                                ROI roi={}, int nthreads=0);
ImageBuf OIIO_API polar_to_complex (const ImageBuf &src,
                                    ROI roi={}, int nthreads=0);
bool OIIO_API polar_to_complex (ImageBuf &dst, const ImageBuf &src,
                                ROI roi={}, int nthreads=0);
/// @}



enum NonFiniteFixMode
{
    NONFINITE_NONE = 0,     ///< Do not alter the pixels (but do count the
                            ///< number of nonfinite pixels in *pixelsFixed,
                            ///< if non-null).
    NONFINITE_BLACK = 1,    ///< Replace non-finite values with 0.0.
    NONFINITE_BOX3 = 2,     ///< Replace non-finite values with the average
                            ///< value of any finite pixels in a 3x3 window.
    NONFINITE_ERROR = 100,  ///< Return false (error), but don't change any
                            ///< values, if any nonfinite values are found.
};

/// `fixNonFinite()` returns in image containing the values of `src` (within
/// the ROI), while repairing any non-finite (NaN/Inf) pixels. If
/// `pixelsFixed` is not nullptr, store in it the number of pixels that
/// contained non-finite value.  It is permissible to operate in-place (with
/// `src` and `dst` referring to the same image).
///
/// How the non-finite values are repaired is specified by one of the `mode`
/// parameter, which is an enum of `NonFiniteFixMode`.
///
/// This function works on all pixel data types, though it's just a copy for
/// images with pixel data types that cannot represent NaN or Inf values.
ImageBuf OIIO_API fixNonFinite (const ImageBuf &src,
                                NonFiniteFixMode mode=NONFINITE_BOX3,
                                int *pixelsFixed = nullptr,
                                ROI roi={}, int nthreads=0);

/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API fixNonFinite (ImageBuf &dst, const ImageBuf &src,
                            NonFiniteFixMode mode=NONFINITE_BOX3,
                            int *pixelsFixed = nullptr,
                            ROI roi={}, int nthreads=0);


/// Copy the specified ROI of `src` and fill any holes (pixels where alpha <
/// 1) with plausible values using a push-pull technique. The `src` image
/// must have an alpha channel.  The `dst` image will end up with a copy of
/// `src`, but will have an alpha of 1.0 everywhere within `roi`, and any
/// place where the alpha of `src` was < 1, `dst` will have a pixel color
/// that is a plausible "filling" of the original alpha hole.
ImageBuf OIIO_API fillholes_pushpull (const ImageBuf &src,
                                      ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API fillholes_pushpull (ImageBuf &dst, const ImageBuf &src,
                                  ROI roi={}, int nthreads=0);


/// Return a median-filtered version of the corresponding region of `src`.
/// The median filter replaces each pixel with the median value underneath
/// the `width` x `height` window surrounding it. If `height` <= 0, it will
/// be set to `width`, making a square window.
///
/// Median filters are good for removing high-frequency detail smaller than
/// the window size (including noise), without blurring edges that are
/// larger than the window size.
ImageBuf OIIO_API median_filter (const ImageBuf &src,
                                 int width = 3, int height = -1,
                                 ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API median_filter (ImageBuf &dst, const ImageBuf &src,
                             int width = 3, int height = -1,
                             ROI roi={}, int nthreads=0);


/// Return a sharpened version of the corresponding region of `src` using
/// the "unsharp mask" technique. Unsharp masking basically works by first
/// blurring the image (low pass filter), subtracting this from the original
/// image, then adding the residual back to the original to emphasize the
/// edges. Roughly speaking,
///
///      dst = src + contrast * thresh(src - blur(src))
///
/// The specific blur can be selected by kernel name and width (for example,
/// "gaussian" is typical). As a special case, "median" is also accepted as
/// the kernel name, in which case a median filter is performed rather than
/// a blurring convolution (Gaussian and other blurs sometimes over-sharpen
/// edges, whereas using the median filter will sharpen compact
/// high-frequency details while not over-sharpening long edges).
/// 
/// The `contrast` is a multiplier on the overall sharpening effect.  The
/// thresholding step causes all differences less than `threshold` to be
/// squashed to zero, which can be useful for suppressing sharpening of
/// low-contrast details (like noise) but allow sharpening of
/// higher-contrast edges.
ImageBuf OIIO_API unsharp_mask (const ImageBuf &src,
                            string_view kernel="gaussian", float width = 3.0f,
                            float contrast = 1.0f, float threshold = 0.0f,
                            ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API unsharp_mask (ImageBuf &dst, const ImageBuf &src,
                            string_view kernel="gaussian", float width = 3.0f,
                            float contrast = 1.0f, float threshold = 0.0f,
                            ROI roi={}, int nthreads=0);


/// Return a dilated version of the corresponding region of `src`. Dilation
/// is defined as the maximum value of all pixels under nonzero values of
/// the structuring element (which is taken to be a width x height square).
/// If height is not set, it will default to be the same as width. Dilation
/// makes bright features wider and more prominent, dark features thinner,
/// and removes small isolated dark spots.
ImageBuf OIIO_API dilate (const ImageBuf &src, int width=3, int height=-1,
                          ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API dilate (ImageBuf &dst, const ImageBuf &src,
                      int width=3, int height=-1, ROI roi={}, int nthreads=0);


/// Return an eroded version of the corresponding region of `src`. Erosion
/// is defined as the minimum value of all pixels under nonzero values of
/// the structuring element (which is taken to be a width x height square).
/// If height is not set, it will default to be the same as width. Erosion
/// makes dark features wider, bright features thinner, and removes small
/// isolated bright spots.
ImageBuf OIIO_API erode (const ImageBuf &src, int width=3, int height=-1,
                         ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API erode (ImageBuf &dst, const ImageBuf &src,
                     int width=3, int height=-1, ROI roi={}, int nthreads=0);



/// @defgroup colorconvert (Color space conversions)
/// @{
///
/// Convert between color spaces
///
/// Return (or copy into `dst`) the pixels of `src` within the ROI, applying
/// a color space transformation. In-place operations (`dst` == `src`) are
/// supported.
///
/// The first three channels are presumed to be the color to be
/// transformed, and the fourth channel (if it exists) is presumed to be
/// alpha. Any additional channels will be simply copied unaltered.
///
/// If OIIO was built with OpenColorIO support enabled, then the
/// transformation may be between any two spaces supported by the active
/// OCIO configuration, or may be a "look" transformation created by
/// `ColorConfig::createLookTransform`.  If OIIO was not built with
/// OpenColorIO support enabled, then the only transformations available are
/// from "sRGB" to "linear" and vice versa.
///
/// @param  fromspace/tospace
///             For the varieties of `colorconvert()` that use named color
///             spaces, these specify the color spaces by name.
/// @param  context_key/context_value
///             For the varieties of `colorconvert()` that use named color
///             spaces, these optionally specify a "key" name/value pair to
///             establish a context (for example, a shot-specific transform).
/// @param  processor
///             For the varieties of `colorconvert()` that have a
///             `processor` parameter, it is a raw `ColorProcessor*` object
///             that implements the color transformation. This is a special
///             object created by a `ColorConfig` (see `OpenImageIO/color.h`
///             for details).
/// @param  unpremult
///             If true, unpremultiply the image (divide the RGB channels by
///             alpha if it exists and is nonzero) before color conversion,
///             then repremult after the after the color conversion. Passing
///             unpremult=false skips this step, which may be desirable if
///             you know that the image is "unassociated alpha" (a.k.a.,
///             "not pre-multiplied colors").
/// @param  colorconfig
///             An optional `ColorConfig*` specifying an OpenColorIO
///             configuration. If not supplied, the default OpenColorIO
///             color configuration found by examining the `$OCIO`
///             environment variable will be used instead.
///

/// Transform between named color spaces, returning an ImageBuf result.
ImageBuf OIIO_API colorconvert (const ImageBuf &src,
                      string_view fromspace, string_view tospace, bool unpremult=true,
                      string_view context_key="", string_view context_value="",
                      const ColorConfig* colorconfig = nullptr,
                      ROI roi={}, int nthreads=0);

/// Transform using a ColorProcessor, returning an ImageBuf result.
ImageBuf OIIO_API colorconvert (const ImageBuf &src,
                                const ColorProcessor *processor,
                                bool unpremult, ROI roi={}, int nthreads=0);
/// Transform between named color spaces, storing reults into an existing ImageBuf.
bool OIIO_API colorconvert (ImageBuf &dst, const ImageBuf &src,
                  string_view fromspace, string_view tospace, bool unpremult=true,
                  string_view context_key="", string_view context_value="",
                  const ColorConfig* colorconfig = nullptr,
                  ROI roi={}, int nthreads=0);

/// Transform using a ColorProcessor, storing reults into an existing ImageBuf.
bool OIIO_API colorconvert (ImageBuf &dst, const ImageBuf &src,
                            const ColorProcessor *processor,
                            bool unpremult, ROI roi={}, int nthreads=0);

/// Apply a color transform in-place to just one color:
/// `color[0..nchannels-1]`.  `nchannels` should either be 3 or 4 (if 4, the
/// last channel is alpha).
bool OIIO_API colorconvert (span<float> color,
                            const ColorProcessor *processor, bool unpremult);

// DEPRECATED(2.1): Less safe version with raw pointer and length.
inline bool colorconvert (float *color, int nchannels,
                          const ColorProcessor *processor, bool unpremult) {
    return colorconvert ({color,nchannels}, processor, unpremult);
}

/// @}


/// Return a copy of the pixels of `src` within the ROI, applying a color
/// transform specified by a 4x4 matrix.  In-place operations
/// (`dst` == `src`) are supported.
///
/// The first three channels are presumed to be the color to be
/// transformed, and the fourth channel (if it exists) is presumed to be
/// alpha. Any additional channels will be simply copied unaltered.
///
/// @param  M
///             A 4x4 matrix. Following Imath conventions, the color is a
///             row vector and the matrix has the "translation" part in
///             elements [12..14] (matching the memory layout of OpenGL or
///             RenderMan), so the math is `color * Matrix` (NOT `M*c`).
/// @param  unpremult
///             If true, unpremultiply the image (divide the RGB channels by
///             alpha if it exists and is nonzero) before color conversion,
///             then repremult after the after the color conversion. Passing
///             unpremult=false skips this step, which may be desirable if
///             you know that the image is "unassociated alpha" (a.k.a.,
///             "not pre-multiplied colors").
///
/// @version 2.1+
///
ImageBuf OIIO_API colormatrixtransform (const ImageBuf &src,
                                        M44fParam M, bool unpremult=true,
                                        ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API colormatrixtransform (ImageBuf &dst, const ImageBuf &src,
                                    M44fParam M, bool unpremult=true,
                                    ROI roi={}, int nthreads=0);


/// Return a copy of the pixels of `src` within the ROI, applying an
/// OpenColorIO "look" transform to the pixel values. In-place operations
/// (`dst` == `src`) are supported.
///
/// The first three channels are presumed to be the color to be
/// transformed, and the fourth channel (if it exists) is presumed to be
/// alpha. Any additional channels will be simply copied unaltered.
///
/// @param  looks
///             The looks to apply (comma-separated).
/// @param  fromspace/tospace
///             For the varieties of `colorconvert()` that use named color
///             spaces, these specify the color spaces by name.
/// @param  unpremult
///             If true, unpremultiply the image (divide the RGB channels by
///             alpha if it exists and is nonzero) before color conversion,
///             then repremult after the after the color conversion. Passing
///             unpremult=false skips this step, which may be desirable if
///             you know that the image is "unassociated alpha" (a.k.a.,
///             "not pre-multiplied colors").
/// @param  inverse
///             If `true`, it will reverse the color transformation and look
///             application.
/// @param  context_key/context_value
///             Optional key/value to establish a context (for example, a
///             shot-specific transform).
/// @param  colorconfig
///             An optional `ColorConfig*` specifying an OpenColorIO
///             configuration. If not supplied, the default OpenColorIO
///             color configuration found by examining the `$OCIO`
///             environment variable will be used instead.
ImageBuf OIIO_API ociolook (const ImageBuf &src, string_view looks,
                            string_view fromspace, string_view tospace,
                            bool unpremult=true, bool inverse=false,
                            string_view context_key="", string_view context_value="",
                            const ColorConfig* colorconfig = nullptr,
                            ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API ociolook (ImageBuf &dst, const ImageBuf &src, string_view looks,
                        string_view fromspace, string_view tospace,
                        bool unpremult=true, bool inverse=false,
                        string_view context_key="", string_view context_value="",
                        const ColorConfig* colorconfig = nullptr,
                        ROI roi={}, int nthreads=0);


/// Return the pixels of `src` within the ROI, applying an OpenColorIO
/// "display" transform to the pixel values. In-place operations
/// (`dst` == `src`) are supported.
///
/// The first three channels are presumed to be the color to be
/// transformed, and the fourth channel (if it exists) is presumed to be
/// alpha. Any additional channels will be simply copied unaltered.
///
/// @param  display
///             The OCIO "display" to apply. If this is `"default"` or the
///             empty string `""`, the default display will be used.
/// @param  view
///             The OCIO "view" to use. If this is `"default"` or the empty
///             string `""`, the default view for this display will be used.
/// @param  fromspace
///             If `fromspace` is not supplied, it will assume that the
///             source color space is whatever is indicated by the source
///             image's metadata or filename, and if that cannot be deduced,
///             it will be assumed to be scene linear.
/// @param  looks
///             The looks to apply (comma-separated). This may be empty,
///             in which case no "look" is used. Note: this parameter value
///             is not used when building against OpenColorIO 2.x.
/// @param  unpremult
///             If true, unpremultiply the image (divide the RGB channels by
///             alpha if it exists and is nonzero) before color conversion,
///             then repremult after the after the color conversion. Passing
///             unpremult=false skips this step, which may be desirable if
///             you know that the image is "unassociated alpha" (a.k.a.,
///             "not pre-multiplied colors").
/// @param  inverse
///             If `true`, it will reverse the color transformation and
///             display application.
/// @param  context_key/context_value
///             Optional key/value to establish a context (for example, a
///             shot-specific transform).
/// @param  colorconfig
///             An optional `ColorConfig*` specifying an OpenColorIO
///             configuration. If not supplied, the default OpenColorIO
///             color configuration found by examining the `$OCIO`
///             environment variable will be used instead.
ImageBuf OIIO_API ociodisplay (const ImageBuf &src,
                               string_view display, string_view view,
                               string_view fromspace="", string_view looks="",
                               bool unpremult=true, bool inverse=false,
                               string_view context_key="", string_view context_value="",
                               const ColorConfig* colorconfig = nullptr,
                               ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API ociodisplay (ImageBuf &dst, const ImageBuf &src,
                           string_view display, string_view view,
                           string_view fromspace="", string_view looks="",
                           bool unpremult=true, bool inverse=false,
                           string_view context_key="", string_view context_value="",
                           const ColorConfig* colorconfig = nullptr,
                           ROI roi={}, int nthreads=0);

#ifndef OIIO_DOXYGEN
// OIIO_DEPRECATED("prefer the kind that takes an `inverse` parameter (2.5)")
ImageBuf OIIO_API ociodisplay (const ImageBuf &src,
                               string_view display, string_view view,
                               string_view fromspace, string_view looks,
                               bool unpremult,
                               string_view context_key, string_view context_value="",
                               const ColorConfig* colorconfig = nullptr,
                               ROI roi={}, int nthreads=0);
// OIIO_DEPRECATED("prefer the kind that takes an `inverse` parameter (2.5)")
bool OIIO_API ociodisplay (ImageBuf &dst, const ImageBuf &src,
                           string_view display, string_view view,
                           string_view fromspace, string_view looks,
                           bool unpremult,
                           string_view context_key, string_view context_value="",
                           const ColorConfig* colorconfig = nullptr,
                           ROI roi={}, int nthreads=0);
#endif


/// Return the pixels of `src` within the ROI, applying an OpenColorIO
/// "file" transform. In-place operations (`dst` == `src`) are supported.
///
/// The first three channels are presumed to be the color to be
/// transformed, and the fourth channel (if it exists) is presumed to be
/// alpha. Any additional channels will be simply copied unaltered.
///
/// @param  name
///             The name of the file containing the transform information.
/// @param  unpremult
///             If true, unpremultiply the image (divide the RGB channels by
///             alpha if it exists and is nonzero) before color conversion,
///             then repremult after the after the color conversion. Passing
///             unpremult=false skips this step, which may be desirable if
///             you know that the image is "unassociated alpha" (a.k.a.,
///             "not pre-multiplied colors").
/// @param  inverse
///             If `true`, it will reverse the color transformation.
/// @param  colorconfig
///             An optional `ColorConfig*` specifying an OpenColorIO
///             configuration. If not supplied, the default OpenColorIO
///             color configuration found by examining the `$OCIO`
///             environment variable will be used instead.
ImageBuf OIIO_API ociofiletransform (const ImageBuf &src,
                                     string_view name,
                                     bool unpremult=true, bool inverse=false,
                                     const ColorConfig* colorconfig = nullptr,
                                     ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API ociofiletransform (ImageBuf &dst, const ImageBuf &src,
                                 string_view name,
                                 bool unpremult=true, bool inverse=false,
                                 const ColorConfig* colorconfig = nullptr,
                                 ROI roi={}, int nthreads=0);


/// @defgroup premult (Premultiply or un-premultiply color by alpha)
/// @{
///
/// Premultiply or un-premultiply color by alpha
///
/// The `unpremult` operation returns (or copies into `dst`) the pixels of
/// `src` within the ROI, and in the process divides all color channels
/// (those not alpha or z) by the alpha value, to "un-premultiply" them.
/// This presumes that the image starts of as "associated alpha" a.k.a.
/// "premultipled," and you are converting to "unassociated alpha." For
/// pixels with alpha == 0, the color values are not modified.
///
/// The `premult` operation returns (or copies into `dst`) the pixels of
/// `src` within the ROI, and in the process multiplies all color channels
/// (those not alpha or z) by the alpha value, to "premultiply" them.  This
/// presumes that the image starts of as "unassociated alpha" a.k.a.
/// "non-premultipled" and converts it to "associated alpha / premultipled."
///
/// The `repremult` operation is like `premult`, but preserves the color
/// values of pixels whose alpha is 0. This is intended for cases where you
/// unpremult, do an operation (such as color transforms), then want to
/// return to associated/premultiplied alpha -- in that case, you want to
/// make sure that "glow" pixels (those with alpha=0 but RGB > 0) are
/// preserved for the round trip, and not crushed to black. This use case is
/// distinct from a simple `premult` that is a one-time conversion from
/// unassociated to associated alpha.
///
/// All three operations are simply a copy if there is no identified alpha
/// channel (and a no-op if `dst` and `src` are the same image).

ImageBuf OIIO_API unpremult (const ImageBuf &src, ROI roi={}, int nthreads=0);
bool OIIO_API unpremult (ImageBuf &dst, const ImageBuf &src,
                         ROI roi={}, int nthreads=0);
ImageBuf OIIO_API premult (const ImageBuf &src, ROI roi={}, int nthreads=0);
bool OIIO_API premult (ImageBuf &dst, const ImageBuf &src,
                       ROI roi={}, int nthreads=0);
ImageBuf OIIO_API repremult (const ImageBuf &src, ROI roi={}, int nthreads=0);
bool OIIO_API repremult (ImageBuf &dst, const ImageBuf &src,
                         ROI roi={}, int nthreads=0);
/// @}



enum MakeTextureMode {
    MakeTxTexture, MakeTxShadow, MakeTxEnvLatl,
    MakeTxEnvLatlFromLightProbe,
    MakeTxBumpWithSlopes,
    _MakeTxLast
};

/// @defgroup make_texture (make_texture -- Turn an image into a texture)
/// @{
///
/// The `make_texture()` function turns an image into a tiled, MIP-mapped,
/// texture file and write it to disk (outputfilename).
///
/// The return value is `true` for success, `false` if an error occurred. If
/// there was an error, any error message will be retrievable via the global
/// `OIIO::geterror()` call (since there is no destination `ImageBuf` in
/// which to store it).
///
/// Named fields in config:
///
///    - format : Data format of the texture file (default: UNKNOWN = same
///               format as the input)
///    - tile_width/tile_height/tile_depth :
///               Preferred tile size (default: 64x64x1)
///
/// Metadata in `config.extra_attribs`
///
///    - `compression` (string) :   Default: "zip"
///    - `fovcot` (float) :         Default: aspect ratio of the image
///                                 resolution.
///    - `planarconfig` (string) :  Default: "separate"
///    - `worldtocamera` (matrix) : World-to-camera matrix of the view.
///    - `worldtoscreen` (matrix) : World-to-screen space matrix of the view.
///    - `worldtoNDC` (matrix) :    World-to-NDC space matrix of the view.
///    - `wrapmodes` (string) :     Default: "black,black"
///    - `handed` (string) :        "left" or "right" reveals the handedness of
///                                 the coordinates for normal maps. ("")
///    - `maketx:verbose` (int) :   How much detail should go to outstream (0).
///    - `maketx:runstats` (int) :  If nonzero, print run stats to outstream (0).
///    - `maketx:resize` (int) :    If nonzero, resize to power of 2. (0)
///    - `maketx:nomipmap` (int) :  If nonzero, only output the top MIP level (0).
///    - `maketx:updatemode` (int) : If nonzero, write new output only if the
///                                  output file doesn't already exist, or is
///                                  older than the input file, or was created
///                                  with different command-line arguments. (0)
///    - `maketx:constant_color_detect` (int) :
///                           If nonzero, detect images that are entirely
///                           one color, and change them to be low
///                           resolution (default: 0).
///    - `maketx:monochrome_detect` (int) :
///                           If nonzero, change RGB images which have
///                           R==G==B everywhere to single-channel
///                           grayscale (default: 0).
///    - `maketx:opaque_detect` (int) :
///                           If nonzero, drop the alpha channel if alpha
///                           is 1.0 in all pixels (default: 0).
///    - `maketx:compute_average` (int) :
///                           If nonzero, compute and store the average
///                           color of the texture (default: 1).
///    - `maketx:unpremult` (int) : If nonzero, unpremultiply color by alpha
///                           before color conversion, then multiply by
///                           alpha after color conversion (default: 0).
///    - `maketx:incolorspace`, `maketx:outcolorspace` (string) :
///                           These two together will apply a color conversion
///                           (with OpenColorIO, if compiled). Default: ""
///    - `maketx:colorconfig` (string) :
///                           Specifies a custom OpenColorIO color config
///                           file. Default: ""
///    - `maketx:checknan` (int) :  If nonzero, will consider it an error if the
///                           input image has any NaN pixels. (0)
///    - `maketx:fixnan` (string) : If set to "black" or "box3", will attempt
///                           to repair any NaN pixels found in the
///                           input image (default: "none").
///    - `maketx:set_full_to_pixels` (int) :
///                           If nonzero, doctors the full/display window
///                           of the texture to be identical to the
///                           pixel/data window and reset the origin
///                           to 0,0 (default: 0).
///    - `maketx:filtername` (string) :
///                           If set, will specify the name of a high-quality
///                           filter to use when resampling for MIPmap
///                           levels. Default: "", use bilinear resampling.
///    - `maketx:highlightcomp` (int) :
///                           If nonzero, performs highlight compensation --
///                           range compression and expansion around
///                           the resize, plus clamping negative pixel
///                           values to zero. This reduces ringing when
///                           using filters with negative lobes on HDR
///                           images.
///    - `maketx:sharpen` (float) : If nonzero, sharpens details when creating
///                           MIPmap levels. The amount is the contrast
///                           metric. The default is 0, meaning no
///                           sharpening.
///    - `maketx:nchannels` (int) : If nonzero, will specify how many channels
///                           the output texture should have, padding with
///                           0 values or dropping channels, if it doesn't
///                           the number of channels in the input.
///                           (default: 0, meaning keep all input channels)
///    - `maketx:channelnames` (string) :
///                           If set, overrides the channel names of the
///                           output image (comma-separated).
///    - `maketx:fileformatname` (string) :
///                           If set, will specify the output file format.
///                           (default: "", meaning infer the format from
///                           the output filename)
///    - `maketx:oiio_options` (int) :
///                           (Deprecated; all are handled by default)
///    - `maketx:prman_options` (int) :
///                           If nonzero, override a whole bunch of settings
///                           as needed to make textures that are
///                           compatible with PRMan. This also enables
///                           prman_metadata. (0)
///    - `maketx:prman_metadata` (int) :
///                           If set, output some metadata that PRMan will
///                           need for its textures. (0)
///    - `maketx:mipimages` (string) :
///                           Semicolon-separated list of alternate images
///                           to be used for individual MIPmap levels,
///                           rather than simply downsizing. (default: "")
///    - `maketx:mipmap_metadata` (int) :
///                           If nonzero, will propagate metadata to every MIP
///                           level. The default (0) only writes metadata to
///                           the highest-resolution level. (0)
///    - `maketx:full_command_line` (string) :
///                           The command or program used to generate this
///                           call, will be embedded in the metadata.
///                           (default: "")
///    - `maketx:ignore_unassoc` (int) :
///                           If nonzero, will disbelieve any evidence that
///                           the input image is unassociated alpha. (0)
///    - `maketx:read_local_MB` (int) :
///                           If nonzero, will read the full input file
///                           locally if it is smaller than this
///                           threshold. Zero causes the system to make a
///                           good guess at a reasonable threshold (e.g. 1
///                           GB). (0)
///    - `maketx:forcefloat` (int) :
///                           Forces a conversion through float data for
///                           the sake of ImageBuf math. (1)
///    - `maketx:hash` (int) :
///                           Compute the sha1 hash of the file in parallel. (1)
///    - `maketx:allow_pixel_shift` (int) :
///                           Allow up to a half pixel shift per mipmap level.
///                           The fastest path may result in a slight shift
///                           in the image, accumulated for each mip level
///                           with an odd resolution. (0)
///    - `maketx:bumpformat` (string) :
///                           For the MakeTxBumpWithSlopes mode, chooses
///                           whether to assume the map is a height map
///                           ("height"), a normal map ("normal"), or
///                           automatically determine it from the number
///                           of channels ("auto", the default).
///    - `maketx:uvslopes_scale` (float) :
///                           If nonzero, when used in MakeTxBumpWithSlopes
///                           mode, this computes derivatives for the
///                           bumpslopes data in UV space rather than in
///                           texel space, and divides them by this scale
///                           factor. The default is 0, disabling the
///                           feature. If you use this feature, a suggested
///                           value is 256.
///    - `maketx:cdf` (int) :
///                           If nonzero, will write a Gaussian CDF and
///                           Inverse Gaussian CDF as per-channel metadata
///                           in the texture, which can be used by shaders
///                           to implement Histogram-Preserving Blending.
///                           This is only useful when the texture being
///                           created is written to an image format that
///                           supports arbitrary metadata (e.g. OpenEXR).
///                           (See Burley, "On Histogram-Preserving Blending
///                           for Randomized Texture Tiling," JCGT 8(4), 2019,
///                           and Heitz/Neyret, "High-Performance By-Example
///                           Noise using a Histogram-Preserving Blending
///                           Operator," ACM SIGGRAPH / Eurographics Symposium
///                           on High-Performance Graphics 2018.) (default: 0)
///    - `maketx:cdfsigma` (float) :
///                           When `maketx:cdf` is active, determines the
///                           CDF sigma (default: 1.0/6).
///    - `maketx:cdfbits` (int) :
///                           When `maketx:cdf` is active, determines the
///                           number of bits to use for the size of the CDF
///                           table. (default: 8, meaning 256 bins)
///
/// @param  mode
///    Describes what type of texture file we are creating and may
///    be one of:
///      - `MakeTxTexture` :  Ordinary 2D texture.
///      - `MakeTxEnvLatl` :  Latitude-longitude environment map
///      - `MakeTxEnvLatlFromLightProbe` :  Latitude-longitude environment map
///                                         constructed from a "light probe"
///                                         image.
///      - `MakeTxBumpWithSlopes` : Bump/displacement map with extra slope
///                                 data channels (6 channels total,
///                                 containing both the height and 1st and
///                                 2nd moments of slope distributions) for
///                                 bump-to-roughness conversion in shaders.
/// @param  outputfilename
///     Name of the file in which to save the resulting texture.
/// @param  config An ImageSpec that contains all the information and
///     special instructions for making the texture.  Anything set in config
///     (format, tile size, or named metadata) will take precedence over
///     whatever is specified by the input file itself.  Additionally, named
///     metadata that starts with "maketx:" will not be output to the file
///     itself, but may contain instructions controlling how the texture is
///     created.  The full list of supported configuration options is listed
///     below.
/// @param  outstream
///     If not `nullptr`, it should point to a stream (for example,
///     `&std::out`, or a pointer to a local `std::stringstream` to capture
///     output), which is where console output and errors will be
///     deposited. Note that error messages will also be retrievable from
///     OIIO::geterror().
///
///

/// Version of make_texture that starts with an ImageBuf.
bool OIIO_API make_texture (MakeTextureMode mode,
                            const ImageBuf &input,
                            string_view outputfilename,
                            const ImageSpec &config,
                            std::ostream *outstream = nullptr);

/// Version of make_texture that starts with a filename and reads the input
/// from that file, rather than being given an ImageBuf directly.
bool OIIO_API make_texture (MakeTextureMode mode,
                            string_view filename,
                            string_view outputfilename,
                            const ImageSpec &config,
                            std::ostream *outstream = nullptr);

/// Version of make_texture that takes multiple filenames (reserved for
/// future expansion, such as assembling several faces into a cube map).
bool OIIO_API make_texture (MakeTextureMode mode,
                            const std::vector<std::string> &filenames,
                            string_view outputfilename,
                            const ImageSpec &config,
                            std::ostream *outstream = nullptr);
/// @}


/// Convert an OpenCV cv::Mat into an ImageBuf, copying the pixels (optionally
/// converting to the pixel data type specified by `convert`, if not UNKNOWN,
/// which means to preserve the original data type if possible).  Return true
/// if ok, false if it was not able to make the conversion from Mat to
/// ImageBuf. Any error messages can be retrieved by calling `geterror()` on
/// the returned ImageBuf. If OpenImageIO was compiled without OpenCV support,
/// this function will return false.
OIIO_API ImageBuf
from_OpenCV (const cv::Mat& mat, TypeDesc convert = TypeUnknown,
             ROI roi={}, int nthreads=0);

/// Construct an OpenCV cv::Mat containing the contents of ImageBuf src, and
/// return true. If it is not possible, or if OpenImageIO was compiled without
/// OpenCV support, then return false. Any error messages can be retrieved by
/// calling OIIO::geterror(). Note that OpenCV only supports up to 4 channels,
/// so >4 channel images will be truncated in the conversion.
OIIO_API bool to_OpenCV (cv::Mat& dst, const ImageBuf& src,
                         ROI roi={}, int nthreads=0);


/// Capture a still image from a designated camera.  If able to do so,
/// store the image in dst and return true.  If there is no such device,
/// or support for camera capture is not available (such as if OpenCV
/// support was not enabled at compile time), return false and do not
/// alter dst.
ImageBuf OIIO_API capture_image (int cameranum = 0,
                                 TypeDesc convert=TypeUnknown);


#ifndef DOXYGEN_SHOULD_SKIP_THIS
// DEPRECATED(1.9):
OIIO_DEPRECATED("use version that returns ImageBuf (1.9)")
inline bool capture_image (ImageBuf &dst, int cameranum = 0,
                           TypeDesc convert=TypeUnknown) {
    dst = capture_image (cameranum, convert);
    return !dst.has_error();
}
#endif  // DOXYGEN_SHOULD_SKIP_THIS


#if defined(__OPENCV_CORE_TYPES_H__) || defined(OPENCV_CORE_TYPES_H)
// These declarations are only visible if the OpenCV headers have already
// been encountered.

#ifndef DOXYGEN_SHOULD_SKIP_THIS
// DEPRECATED(2.0). The OpenCV 1.x era IplImage-based functions should be
// avoided, giving preference to from_OpenCV.
ImageBuf OIIO_API from_IplImage (const IplImage *ipl,
                                 TypeDesc convert=TypeUnknown);
// DEPRECATED(1.9):
OIIO_DEPRECATED("use from_OpenCV (1.9)")
inline bool from_IplImage (ImageBuf &dst, const IplImage *ipl,
                           TypeDesc convert=TypeUnknown) {
    dst = from_IplImage (ipl, convert);
    return ! dst.has_error();
}

// DEPRECATED(2.0). The OpenCV 1.x era IplImage-based functions should be
// avoided, giving preference to from_OpenCV.
OIIO_DEPRECATED("use from_OpenCV (2.9)")
OIIO_API IplImage* to_IplImage (const ImageBuf &src);
#endif  // DOXYGEN_SHOULD_SKIP_THIS

#endif



/// Return the "deep" equivalent of the "flat" input `src`. Turning a flat
/// image into a deep one means:
///
/// * If the `src` image has a "Z" channel: if the source pixel's Z channel
///   value is not infinite, the corresponding pixel of `dst` will get a
///   single depth sample that copies the data from the source pixel;
///   otherwise, dst will get an empty pixel. In other words, infinitely far
///   pixels will not turn into deep samples.
///
/// * If the `src` image lacks a "Z" channel: if any of the source pixel's
///   channel values are nonzero, the corresponding pixel of `dst` will get
///   a single depth sample that copies the data from the source pixel and
///   uses the zvalue parameter for the depth; otherwise, if all source
///   channels in that pixel are zero, the destination pixel will get no
///   depth samples.
///
/// If `src` is already a deep image, it will just copy pixel values from
/// `src`.
ImageBuf OIIO_API deepen (const ImageBuf &src, float zvalue = 1.0f,
                          ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API deepen (ImageBuf &dst, const ImageBuf &src, float zvalue = 1.0f,
                      ROI roi={}, int nthreads=0);


/// Return the "flattened" composite of deep image `src`. That is, it
/// converts a deep image to a simple flat image by front-to- back
/// compositing the samples within each pixel.  If `src` is already a
/// non-deep/flat image, it will just copy pixel values from `src` to `dst`.
/// If `dst` is not already an initialized ImageBuf, it will be sized to
/// match `src` (but made non-deep).
ImageBuf OIIO_API flatten (const ImageBuf &src, ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API flatten (ImageBuf &dst, const ImageBuf &src,
                       ROI roi={}, int nthreads=0);


/// Return the deep merge of the samples of deep images `A` and `B`,
/// overwriting any existing samples of `dst` in the ROI. If
/// `occlusion_cull` is true, any samples occluded by an opaque sample will
/// be deleted.
ImageBuf OIIO_API deep_merge (const ImageBuf &A, const ImageBuf &B,
                              bool occlusion_cull = true,
                              ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API deep_merge (ImageBuf &dst, const ImageBuf &A,
                          const ImageBuf &B, bool occlusion_cull = true,
                          ROI roi={}, int nthreads=0);


/// Return the samples of deep image `src` that are closer than the opaque
/// frontier of deep image holdout, returning true upon success and false
/// for any failures. Samples of `src` that are farther than the first
/// opaque sample of holdout (for the corresponding pixel)will not be copied
/// to `dst`. Image holdout is only used as the depth threshold; no sample
/// values from holdout are themselves copied to `dst`.
ImageBuf OIIO_API deep_holdout (const ImageBuf &src, const ImageBuf &holdout,
                            ROI roi={}, int nthreads=0);
/// Write to an existing image `dst` (allocating if it is uninitialized).
bool OIIO_API deep_holdout (ImageBuf &dst, const ImageBuf &src,
                            const ImageBuf &holdout,
                            ROI roi={}, int nthreads=0);




///////////////////////////////////////////////////////////////////////
// DEPRECATED(1.9): These are all functions that take raw pointers,
// which we are deprecating as of 1.9, replaced by new versions that
// take span<> for length safety.

#ifndef DOXYGEN_SHOULD_SKIP_THIS

//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool fill (ImageBuf &dst, const float *values,
                  ROI roi={}, int nthreads=0) {
    int nc (roi.defined() ? roi.nchannels() : dst.nchannels());
    return fill (dst, {values, nc}, roi, nthreads);
}
//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool fill (ImageBuf &dst, const float *top, const float *bottom,
                  ROI roi={}, int nthreads=0) {
    int nc (roi.defined() ? roi.nchannels() : dst.nchannels());
    return fill (dst, {top, nc}, {bottom, nc}, roi, nthreads);
}
//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool fill (ImageBuf &dst, const float *topleft, const float *topright,
                  const float *bottomleft, const float *bottomright,
                  ROI roi={}, int nthreads=0) {
    int nc (roi.defined() ? roi.nchannels() : dst.nchannels());
    return fill (dst, {topleft, nc}, {topright, nc}, {bottomleft, nc},
                 {bottomright, nc}, roi, nthreads);
}

//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool checker (ImageBuf &dst, int width, int height, int depth,
                     const float *color1, const float *color2,
                     int xoffset=0, int yoffset=0, int zoffset=0,
                     ROI roi={}, int nthreads=0) {
    int nc (roi.defined() ? roi.nchannels() : dst.nchannels());
    return checker (dst, width, height, depth, {color1,nc}, {color2,nc},
                    xoffset, yoffset, zoffset, roi, nthreads);
}

//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool add (ImageBuf &dst, const ImageBuf &A, const float *B,
                 ROI roi={}, int nthreads=0) {
    return add (dst, A, {B,A.nchannels()}, roi, nthreads);
}
//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool sub (ImageBuf &dst, const ImageBuf &A, const float *B,
                 ROI roi={}, int nthreads=0) {
    return sub (dst, A, {B,A.nchannels()}, roi, nthreads);
}
//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool absdiff (ImageBuf &dst, const ImageBuf &A, const float *B,
                     ROI roi={}, int nthreads=0) {
    return absdiff (dst, A, cspan<float>(B,A.nchannels()), roi, nthreads);
}
//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool mul (ImageBuf &dst, const ImageBuf &A, const float *B,
                 ROI roi={}, int nthreads=0) {
    return mul (dst, A, {B, A.nchannels()}, roi, nthreads);
}
//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool div (ImageBuf &dst, const ImageBuf &A, const float *B,
                 ROI roi={}, int nthreads=0) {
    return div (dst, A, {B, A.nchannels()}, roi, nthreads);
}
//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool mad (ImageBuf &dst, const ImageBuf &A, const float *B,
                 const ImageBuf &C, ROI roi={}, int nthreads=0) {
    return mad (dst, A, {B, A.nchannels()}, C, roi, nthreads);
}
//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool mad (ImageBuf &dst, const ImageBuf &A, const ImageBuf &B,
                 const float *C, ROI roi={}, int nthreads=0) {
    return mad (dst, A, C, B, roi, nthreads);
}
//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool mad (ImageBuf &dst, const ImageBuf &A, const float *B,
                 const float *C, ROI roi={}, int nthreads=0) {
    return mad (dst, A, {B, A.nchannels()}, {C, A.nchannels()}, roi, nthreads);
}

//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool pow (ImageBuf &dst, const ImageBuf &A, const float *B,
                 ROI roi={}, int nthreads=0) {
    return pow (dst, A, {B, A.nchannels()}, roi, nthreads);
}

//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool channel_sum (ImageBuf &dst, const ImageBuf &src,
                         const float *weights=nullptr, ROI roi={},
                         int nthreads=0) {
    return channel_sum (dst, src, {weights, src.nchannels()},
                        roi, nthreads);
}

//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool channels (ImageBuf &dst, const ImageBuf &src,
                      int nchannels, const int *channelorder,
                      const float *channelvalues=nullptr,
                      const std::string *newchannelnames=nullptr,
                      bool shuffle_channel_names=false, int nthreads=0) {
    return channels (dst, src, nchannels,
                     { channelorder, channelorder?nchannels:0 },
                     { channelvalues, channelvalues?nchannels:0 },
                     { newchannelnames, newchannelnames?nchannels:0},
                     shuffle_channel_names, nthreads);
}

//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool clamp (ImageBuf &dst, const ImageBuf &src,
                   const float *min=nullptr, const float *max=nullptr,
                   bool clampalpha01 = false,
                   ROI roi={}, int nthreads=0) {
    return clamp (dst, src, { min, min ? src.nchannels() : 0 },
                  { max, max ? src.nchannels() : 0 }, clampalpha01,
                  roi, nthreads);
}

//OIIO_DEPRECATED("use version that takes span<> instead of raw pointer (2.0)")
inline bool isConstantColor (const ImageBuf &src, float *color,
                             ROI roi={}, int nthreads=0) {
    int nc = roi.defined() ? std::min(roi.chend,src.nchannels()) : src.nchannels();
    return isConstantColor (src, {color, color ? nc : 0}, roi, nthreads);
}

//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool color_count (const ImageBuf &src, imagesize_t *count,
                         int ncolors, const float *color,
                         const float *eps=nullptr,
                         ROI roi={}, int nthreads=0) {
    return color_count (src, count, ncolors,
                        { color, ncolors*src.nchannels() },
                        eps ? cspan<float>(eps,src.nchannels()) : cspan<float>(),
                        roi, nthreads);
}

//OIIO_DEPRECATED("use version that takes span<> instead of raw pointer (2.0)")
inline bool color_range_check (const ImageBuf &src, imagesize_t *lowcount,
                               imagesize_t *highcount, imagesize_t *inrangecount,
                               const float *low, const float *high,
                               ROI roi={}, int nthreads=0) {
    return color_range_check (src, lowcount, highcount, inrangecount,
                              {low,src.nchannels()}, {high,src.nchannels()},
                              roi, nthreads);
}

//OIIO_DEPRECATED("use version that takes cspan<> instead of raw pointer (2.0)")
inline bool render_text (ImageBuf &dst, int x, int y, string_view text,
                         int fontsize, string_view fontname,
                         const float *textcolor) {
    return render_text (dst, x, y, text, fontsize, fontname,
                        {textcolor, textcolor?dst.nchannels():0});
}

#endif  // DOXYGEN_SHOULD_SKIP_THIS

///////////////////////////////////////////////////////////////////////

}  // end namespace ImageBufAlgo

OIIO_NAMESPACE_END