OpenImageIO Release 2.0.3 (first stable 2.0 release)

Release 2.0 (Dec 7, 2018) -- compared to 1.8.x

New minimum dependencies:

• On Windows compiling with MSVS, the new minimum version is MSVS 2015.

Major new features and improvements:

• ImageInput and ImageOutput static create() and open() methods now return
  unique_ptr rather than raw pointers. #1934, #1945 (1.9.3).
• ImageInput improvements to thread safety and concurrency, including some
  new API calls (see "Public API changes" section below).
• ImageBufAlgo overhaul (both C++ and Python): Add IBA functions that
  return image results directly rather than passing ImageBuf references
  as parameters for output (the old kind of calls still exist, too, and
  have their uses). Also in C++, change all IBA functions that took raw
  pointers to per-channel colors into span<> for safety. #1961 (1.9.4)
• For some readers and writers, an "IOProxy" can be passed that customizes
  the I/O methods. An important use of this is to write an image "file"
  to memory or to read an image "file" from a memory, rather than disk.
  Currently, OpenEXR supports this for both reading and writing, and PNG
  supports it for writing. You specify a pointer to the proxy via the
  configuration option "oiio:ioproxy". #1931 (1.9.3)
• New Image Format support:
  • OpenVDB file read (as volume images or accessing via texture3d()).
    #2010,2018 (1.9.4)
  • "null" images -- null reader just returns black (or constant colored)
    pixels, null writer just returns. This can be used for benchmarking
    (to eliminate all actual file I/O time), "dry run" where you want to
    test without creating output files. #1778 (1.9.0), #2042 (1.9.4)
• TIFF I/O of multiple scanlines or tiles at once (or whole images, as is
  typical use case for oiiotool and maketx) is sped up by a large factor
  on modern multicore systems. We've seen 10x or more faster oiiotool
  performance for uint8 and uint16 TIFF files using "zip" (deflate)
  compression, on modern 12-16 core machines. #1853 (1.9.2)
• Major refactor of Exif metadata handling, including much more complete
  metadata support for RAW formats and support of camera "maker notes"
  for Canon cameras. #1774 (1.9.0)
• New maketx option --bumpslopes specifically for converting bump maps,
  saves additional channels containing slope distribution moments that can
  be used in shaders for "bump to roughness" calculations. #1810,#1913,2005
  (1.9.2), #2044 (1.9.4)
• An official FindOpenImageIO.cmake that we invite you to use in other
  cmake-based projects that needs to find OIIO. #2027 (1.9.4)

Public API changes:

• Python binding overhaul
  The Python bindings have been reimplemented with
  pybind11, no longer with Boost.Python.
  #1801 (1.9.1)
  In the process (partly due to what's easy or hard in pybind11, but partly
  just because it caused us to revisit the python APIs), there are some minor
  API changes, some of which are breaking! To wit:
  • All of the functions that are passed or return blocks of pixels
    (such as ImageInput.read_image()) now use Numpy ndarray objects
    indexed as [y][x][channel] (no longer using old-style Python
    array.array and flattened to 1D).
  • Specilized enum type ImageInput.OpenMode has been replaced by string
    parameters, so for example, old ImageInput.open(filename, ImageInput.Create)
    is now ImageInput.open (filename, "Create")
  • Any function that previously took a parameter of type TypeDesc
    or TypeDesc.BASETYPE now will accept a string that signifies the
    type. For example, ImageBuf.set_write_format("float") is now a
    synonym for ImageBuf.set_write_format(oiio.TypeDesc(oiio.FLOAT)).
  • For several color conversion functions, parameter names were changed
    from "from" to "fromspace" and "to" to "tospace" to avoid a clash with
    the Python reserved word from. #2084
• ImageInput API changes for thread safety and statelessness #1927 (1.9.2)
  • seek_subimage() no longer takes an ImageSpec&, to avoid the obligatory
    copy. (If the copy is desired, just call spec() to get it afterwards.)
  • All of the read_*() methods now have varieties that take arguments
    specifying the subimage and mip level. The read_native_*() methods
    supplied by ImageInput subclass implementations now ONLY come in the
    variety that takes a subimage and miplevel.
  • All of the read_*() calls that take subimage/miplevel explicitly are
    guaranteed to be stateless and thread-safe against each other (it's not
    necessary to call seek_subimage first, nor to have to lock a mutex to
    ensure that another thread doesn't change the subimage before you get a
    chance to call read). For back-compatibility, there are still versions
    that don't take subimage/miplevel, require a prior call to seek_subimge,
    and are thus not considered thread-safe.
  • New methods spec(subimage,miplevel) and spec_dimensions(s,m)
    let you retrieve a copy of the ImageSpec for a given subimage and
    MIP level (thread-safe, and without needing a prior seek_subimage)
    call. Note that to be stateless and thread-safe, these return a COPY
    of the spec, rather than the reference returned by the stateful
    spec() call that has no arguments and requires a prior seek_subimage.
    However, spec_dimensions() does not copy the channel names or the
    arbitrary metadata, and is thus very inexpensive if the only thing
    you need from the spec copy is the image dimensions and channel
    formats.
• ImageInput and ImageOutput create/open changes
  • The static create() and open() methods have been changed so that
    instead of returning an ImageInput * (or ImageOutput *) and
    requiring the caller to correctly manage that resource and eventually
    destroy it, now they return a unique_ptr that automatically deletes
    when it leaves scope. In the process we also clean up some edge cases
    on Windows where it was possible for ImageInput/ImageOutput to have
    been allocated in one DLL's heap but freed in a different DLL's heap,
    which could cause subtle heap corruption problems. #1934,#1945 (1.9.3).
• ImageBuf
  • New method set_origin() changes the pixel data window origin.
    #1949 (1.9.4)
  • Assignment (operator=) is now enabled for ImageBuf, both the copying
    and moving variety. Also, an explicit copy() method has been added
    that returns a full copy of the ImageBuf. #1952 (1.9.4)
  • write() method has had its arguments changed and now takes an optional
    TypeDesc that lets you specify a requested data type when writing the
    output file, rather than requiring a previous and separate call to
    set_write_format(). The old call signature of write() still exists,
    but it will be considered deprecated in the future. #1953 (1.9.4)
• ImageBufAlgo
  • In C++, functions that take raw pointers for per-channel constant
    values or results are deprecated, in favor of new versions that
    heavily rely on span<> to safely pass array references and their
    lengths. #1961 (1.9.4)
  • In both C++ and Python, every IBA function that takes a parameter
    giving an ImageBuf destination reference for results have an additional
    variant that directly returns an ImageBuf result. This makes much
    cleaner, more readable code, in cases where it's not necessary to
    write partial results into an existing IB. #1961 (1.9.4)
  • In C++, many IBA functions that came in multiple versions for whether
    certain parameters could be an image, a per-channel constant, or a
    single constant, have been replaced by a single version that takes
    a new parameter-passing helper class, Image_or_Const that will match
    against any of those choices. (No changes are necessary for calling
    programs, but it makes the header and documentation a lot simpler.)
    #1961 (1.9.4)
  • IBA compare(), computePixelStats(), and histogram() now directly
    return their result structures, intead of requiring the passing of
    a destination reference. #1961 (1.9.4)
  • New IBA::fit() resizes and image to just fit in the given size, but
    preserve its aspect ratio (padding with black as necessary). It's just
    like what oiiotool --fit has always done, but now you can call it
    directly from C++ or Python. #1993 (1.9.4)
  • New contrast_remap() allows flexible linear or sigmoidal contrast
    remapping. #2043 (1.9.4)
  • ImageBufAlgo::colorconvert and various ocio transformations have
    changed the default value of their unpremult parameter from false
    to true, reflecting the fact that we believe this is almost always
    the more correct choice. Also, if their input image is clearly marked
    as having unasociated alpha already, they will not bracket the color
    conversion with the requested unpremult/premult. #1864 (1.9.2)
  • Updated the OpenCV interoperability with new functions to_OpenCV (make
    an ImageBuf out of a cv::Mat) and from_OpenCV (fill in a cv::Mat with
    the contents of an ImageBuf). Deprecated the old from_IplImage and
    to_IplImage, which are very OpenCV-1.x-centric. (2.0.2)
• ImageCache/TextureSystem:
  • ImageCache and TextureSystem now have close(filename) and
    close_all() methods, which for one file or all files will close the
    files and release any open file handles (also unlocking write access
    to those files on Windows), but without invalidating anything it knows
    about the ImageSpec or any pixel tiles already read from the files, as
    would happen with a call to the much more drastic invalidate() or
    invalidate_all(). #1950 (1.9.4)
  • TextureSystem::create() has an additional optional argument that
    allows the caller to pass an existing app-owned custom ImageCache.
    #2019 (1.9.4)
  • New TextureSystem::imagecache() method returns a blind, non-owning
    pointer to the underlying ImageCache of that TS. #2019 (1.9.4)
  • ImageCache: extended add_tile() with an optional copy parameter
    (which defaults to true), which when set to false will make a tile
    that references an app buffer without allocating, copying, and owning
    the memory. In short, this makes it possible to reference existing
    memory holding an image array, as if it were a texture. #2012 (1.9.4)
  • ImageCache::add_file() extended with an optional replace parameter
    (default: false), that if true, will replace the tile and invalidate
    the old one. #2021 (1.9.4)
• Changes to string formatting: #2076 (2.0.1)
  • New Strutil::sprintf() and ustring::sprintf() functions are for
    printf-style formatted errors and warnings. You are encouraged to
    change your existing format() calls to sprintf(), since the
    original format may in a later version (2.1?) switch to Python-style
    formatting commands, but sprintf will continue to reliably use
    C printf style notation.
  • In ImageInput, ImageOutput, ImageBuf, and ErrorHandler, new errorf()
    and warningf() methods similarly provide printf-style formatted
    errors and warnings. The old error()/warning() calls will someday
    (maybe 2.1?) switch to Python-style formatting commands, but
    errorf will continue to reliably use C printf style notation.
• ColorConfig changes: ColorConfig methods now return shared pointers to
  ColorProcessor rather than raw pointers. It is therefore no longer
  required to make an explicit delete call. Created ColorProcessor objects
  are now internally cached, so asking for the same color transformation
  multiple times is no longer expensive. The ColorProcessor interface is
  now in color.h and can be directly used to perform transformations on
  individual colors (previously it was just an opaque pointer and could
  only be used to pass into certain IBA functions). The color space names
  "rgb" and "default" are now understood to be synonyms for the default
  "linear" color space. #1788 (1.9.0)
• Remove long-deprecated API calls:
  • ImageBuf::get_pixels/get_pixel_channels varieties deprecated since 1.6.
  • ImageBuf::set_deep_value_uint, deprecated since 1.7.
  • ImageBuf::deep_alloc, deprecated since 1.7.
  • ImageBufAlgo::colorconvert variety deprecated since 1.7.
  • ImageCache::clear, deprecated since 1.7.
  • ImageCache::add_tile variety deprecated since 1.6.
• ROI new methods: contains() #1874, #1878 (1.9.2)
• ImageBufAlgo::pixeladdr() now takes an additional optional parameter,
  the channel number. #1880 (1.9.2)
• Global OIIO attribute "log_times" (which defaults to 0 but can be overridden
  by setting the OPENIMAGEIO_LOG_TIMES environment variable), when nonzero,
  instruments ImageBufAlgo functions to record the number of times they are
  called and how much time they take to execute. A report of these times
  can be retrieved as a string as the "timing_report" attribute, or it will
  be printed to stdout automatically if the value of log_times is 2 or more
  at the time that the application exits. #1885 (1.9.2)
• Moved the definition of ROI from imagebuf.h to imageio.h and make
  most of the methods constexpr. #1906 (1.9.2)
• Rename/move of array_view to span. Deprecated array_view and moved
  array_view.h contents to span.h. You should change array_view<T>
  to span<T> and array_view<const T> to cspan<T>. #1956,2062 (1.9.4)
• ustring: removed operator int() that allowed simple int casting such as:

      ustring u, v;
      if (u || !v) { ... }
  

  This was error-prone, neither std::string nor std::string_view had the
  equivalent, so we are removing it. The preferred idiom is:

      if (!u.empty() || v.empty()) { ... }
  

Performance improvements:

• ImageBufAlgo::computePixelStats is now multithreaded and should improve by
  a large factor when running on a machine with many cores. This is
  particularly noticable for maketx. #1852 (1.9.2)
• Color conversions are sped up by 50% for 4 channel float images, about
  30% for other combinations of channels or data formats. #1868 (1.9.2)
• ImageBuf::get_pixels() sped up by around 3x for the common case of the
  image being fully in memory (the slower path is now only used for
  ImageCache-based images). #1872 (1.9.2)
• ImageBufAlgo::copy() and crop() sped up for in-memory buffers, by about
  35-45% when copying between buffers of the same type, 2-4x when copying
  between buffers of different data types. #1877 (1.9.2)
• ImageBufAlgo::over() when both buffers are in-memory, float, 4-channels,
  sped up by about 2x. #1879 (1.9.2).
• ImageBufAlgo::fill() of a constant color sped up by 1.5-2.5x (depending
  on the data type involved). #1886 (1.9.2)

Fixes and feature enhancements:

• oiiotool
  • --help prints important usage tips that explain command parsing,
    syntax of optional modifiers, and the path to PDF docs. #1811 (1.9.2)
  • --colormap has new maps "inferno", "magma", "plasma", "viridis",
    which are perceptually uniform, monotonically increasing luminance,
    look good converted to greyscale, and usable by people with color
    blindness. #1820 (1.9.2)
  • oiiotool no longer enables autotile by default. #1856 (1.9.2)
  • --colorconvert, --tocolorspace, and all of the --ocio commands
    now take an optional modifier :unpremult=1 which causes the color
    conversion to be internally bracketed by unpremult/premult steps (if
    the image has alpha and is not already marked as having unassociated
    alpha). You should therefore prefer --colorconvert:unpremult=1 from to
    rather than the more complex --unpremult --colorconvert from to -premult.
    #1864 (1.9.2)
  • --autocc will also cause unpremult/premult to bracket any color
    transformations it does automatically for read and write (if the image
    has alpha and does not appear to already be unassociated). #1864 (1.9.2)
  • --help prints the name of the OCIO color config file. #1869 (1.9.2)
  • Frame sequence wildcard improvements: fix handling of negative frame
    numbers and ranges, also the --frames command line option is not
    enough to trigger a loop over those frame numbers, even if no other
    arguments appear to have wildcard structure. #1894 (1.8.10/1.9.2)
  • --info -v now prints metadata in sorted order, making it easier to
    spot the existance of particular metadata. #1982 (1.9.4)
  • --no-autopremult fixed, it wasn't working properly for cases that
    were read directly rather than backed by ImageCache. #1984 (1.9.4)
  • New --contrast allows for contrast remapping (linear or sigmoidal).
    #2043 (1.9.4)
  • Improved logic for propagating the pixel data format through
    multiple operations, especially for files with multiple subimages.
    #1769 (1.9.0/1.8.6)
  • Outputs are now written to temporary files, then atomically moved
    to the specified filename at the end. This makes it safe for oiiotool
    to "overwrite" a file (i.e. oiiotool in.tif ... -o out.tif) without
    problematic situations where the file is truncated or overwritten
    before the reading is complete. #1797 (1.8.7/1.9.1)
  • Fixed problem with reading half files where very small (denormalized)
    half values could get squashed to float 0.0 instead of having their
    values preserved, if certain old versions of libopenjpeg were being
    used (because they set a CPU flag strangely upon library load and then
    never changed it back, this is a libopenjpeg bug that has since been
    fixed). #2048 (2.0)
  • -d chan=type logic fixed for certain cases of specifying the data
    types of individual channels. #2061 (2.0beta2)
  • Expression evaluation: metadata names can now be enclosed in single
    or double quotes if they don't follow "C" identifier naming conventions.
    For example, {TOP.'foo/bar'} retrieves metadata called "foo/bar"
    rather than trying to retrieve "foo" and divide by bar. #2068 (2.0beta2)
  • Expression evaluation: When retrieving metadata, timecode data will be
    expressed properly as a string ("00:00:00:00"). #2068 (2.0beta2)
• ImageBufAlgo:
  • color_map() supports new maps "inferno", "magma", "plasma",
    "viridis". #1820 (1.9.2)
  • Across many functions, improve channel logic when combining an image
    with alpha with another image without alpha. #1827 (1.9.2)
  • mad() now takes an img*color+img variety. (Previously it
    supported img*img+img and img*color+color.) #1866 (1.9.2)
  • New fit() is like resize but fits inside a specified window size,
    while preserving the aspect ratio of the image appearance. #1993.
  • New contrast_remap() allows flexible linear or sigmoidal contrast
    remapping. #2043 (1.9.4)
  • channel_append() is no longer limited to requiring the two input
    images to have the same pixel data type. #2022 (1.9.4)
  • isConstantColor(), isConstantChannel(), and isMonochrome() have
    added an optional threshold parameter that allows you to compute
    whether the image is constant or monochrome within a non-zero
    tolerance (the default is still 0.0, meaning checking for an exact
    match). #2049 (2.0.0)
  • IBA::ociodisplay() has better behavior when its "fromspace" parameter
    is left blank -- instead of assuming "linear" (as a space name), it
    assumes it's whatever space in your OCIO color config has the "linear"
    role. #2083 (1.8.17/2.0.1)
• ImageBuf:
  • Bug fixed in IB::copy() of rare types. #1829 (1.9.2)
  • write() automatically tells the ImageCache to 'invalidate' the file
    being written, so cached images will not retain the prior version of
    the files. #1916 (1.9.2)
  • Bug fix to ImageBuf::contains_roi() method -- it erroneously always
    returned true. #1997 (1.8.14/1.9.4)
• ImageCache/TextureSystem/maketx:
  • Improved stats on how long we wait for ImageInput mutexes.
    #1779 (1.9.0/1.8.6)
  • Improved performance of IC/TS tile and file caches under heavy
    contention from many threads. #1780 (1.9.0)
  • Increased the default max_tile_channels limit from 5 to 6.
    #1803 (1.9.1)
  • maketx: improved image hashing to avoid some (extremely rare) possible
    hash collisions. #1819 (1.9.2)
  • IC/TS performance improvements by changing the underlying hash table
    implementation. #1823,1824,1825,1826,1830 (1.9.2)
  • texture()/texture3d(): when requesting a nonexistant "subimage",
    return the fill color, like we do when requesting nonexistant channels
    (rather than nondeterministically simply not filling in the result).
    #1917 (1.9.2)
  • Relying on some changes to the ImageInput API, there is now much less
    thread locking to protect the underlying ImageInputs, and this should
    improve texture and image cache performance when many threads need
    to read tiles from the same file. #1927 (1.9.2)
  • get_image_info()/get_texture_info() is now more flexible about
    retrieving arrays vs aggregates, in cases where the total number of
    elements is correct. #1968 (1.9.4)
  • Fix uninitialized read within the texture system (only affected
    statistics, never gave wrong texture results). #2000 (1.9.4)
  • texture3d() transforms lookup points from world into local space
    if the file has a "worldtolocal" metadata giving a 4x4 matrix. #2009
    (1.9.4)
  • Fix minor texture filtering bug where widely disparate "sblur" and
    "tblur" values could in some circumstances lead to incorrect texture
    filter estimation. #2052 (2.0.0)
  • ImageCache::invalidate(filename) did not properly invalidate the
    "fingerprint" is used to detect duplicate files. #2081 (1.8.17/2.0.1)
• iv:
  • Fix (especially on OSX) for various ways it has been broken since the
    shift to Qt5. #1946 (1.8.12, 1.9.4)
  • New optin --no-autopremult works like oiiotool, causes images with
    unassociated alpha to not be automatically premultiplied by alpha
    as they are read in. #1984 (1.9.4)
• All string->numeric parsing and numeric->string formatting is now
  locale-independent and always uses '.' as decimal marker. #1796 (1.9.0)
• Python Imagebuf.get_pixels and set_pixels bugs fixed, in the varieties
  that take an ROI to describe the region. #1802 (1.9.2)
• Python: Implement missing ImageOutput.open() call variety for declaring
  multiple subimages. #2074 (2.0.1)
• More robust parsing of XMP metadata for unknown metadata names.
  #1816 (1.9.2/1.8.7)
• Fix ImageSpec constructor from an ROI, display/"full" window did not get
  the right default origin. #1997 (1.8.14/1.9.4)
• ImageSpec::erase_attribute() fix bug where it got case-sensitivity of the
  search backwards when built using std::regex rather than boost::regex.
  #2003 (1.8.14/1.9.4)
• DPX:
  • Better catching of write errors, including filling the disk while in
    the process of writing a DPX file. #2072 (2.0.1)
• Field3d:
  • Prevent crashes when open fails. #1848 (1.9.2/1.8.8)
  • Fix potential mutex deadlock. #1972 (1.9.4)
• GIF:
  • Fix crash when reading GIF with comment extension but no comment data.
    #2001 (1.8.14/1.9.4)
• JPEG:
  • When writing, be robust to accidentally setting the "density" metadata
    to values larger than JPEG's 16 bit integer field will accommodate.
    #2002 (1.8.14/1.9.4)
  • Better detection and reporting of error conditions while reading
    corrupt JPEG files. #2073 (2.0.1)
• OpenEXR:
  • Gracefully detect and reject files with subsampled channels,
    which is a rarely-to-never-used OpenEXR feature that we don't support
    properly. #1849 (1.9.2/1.8.8)
  • Improved handling of UTF-8 filenames on Windows. #1941 (1.9.3, 1.8.12,
    1.7.19)
• PNG:
  • Fix redundant png_write_end call. #1910 (1.9.2)
• PSD:
  • Fix parse issue of layer mask data. #1777 (1.9.2)
• RAW:
  • Add "raw:HighlightMode" configuration hint to control libraw's
    handling of highlight mode processing. #1851
  • Important bug fix when dealing with rotated (and vertical) images,
    which were not being re-oriented properly and could get strangely
    scrambled. #1854 (1.9.2/1.8.9)
  • Major rewrite of the way makernotes and camera-specific metadata are
    handled, resulting in much more (and more accurate) reporting of
    camera metadata. #1985 (1.9.4)
  • The "oiio:ColorSpace" metadata is now set correctly when reading
    raw DSLR images. And we deprecate the old "raw:ColorSpace" metadata,
    which is useless. #2016 (1.9.4)
  • Add "raw:aber" configuration hint to control libraw's adjustments for
    chromatic aberration. This data is of type "float[2]", the first value
    is the scale factor for red, the second for blue, and both should be
    very close to 1.0. #2030 (1.9.4)
• TIFF:
  • Improve performance of TIFF scanline output. #1833 (1.9.2)
  • Bug fix: read_tile() and read_tiles() input of un-premultiplied tiles
    botched the "shape" of the tile data array. #1907 (1.9.2/1.8.10)
  • Improvement in speed of reading headers (by removing redundant call
    to TIFFSetDirectory). #1922 (1.9.2)
  • When config option "oiio:UnassociatedAlpha" is nonzero (or not set
    -- which is the default), therefore enabling automatic premultiplication
    by alpha for any unassociated alpha files, it will set the metadata
    "tiff:UnassociatedAlpha" to indicate that the original file was
    unassociated. #1984 (1.9.4)
  • Bug fixes for TIFF reads of images with unassociated alpha -- there
    were some edge cases where they pixels failed to automatically
    premultiply upon read. #2032 (1.9.4)
• zfile: more careful gzopen on Windows that could crash when given bogus
  filename. #1839,2070 (1.9.2/1.8.8/2.0.1)
• Windows fix: Safer thread pool destruction on. #2038 (1.9.4)

Build/test system improvements and platform ports:

• Fixes for Windows build. #1793, #1794 (1.9.0/1.8.6), #2025 (1.9.4)
• Fix build bug where if the makefile wrapper got CODECOV=0, it would
  force a "Debug" build (required for code coverage tests) even though code
  coverage is instructed to be off. (It would be fine if you didn't specify
  CODECOV at all.) #1792 (1.9.0/1.8.6)
• Build: Fix broken build when Freetype was not found or disabled. #1800
  (1.8.6/1.9.1)
• Build: Boost.Python is no longer a dependency, but pybind11 is. If
  not found on the system, it will be automatically downloaded. #1801, #2031
  (1.9.1)
• Time for a multi-core build of OIIO is reduced by 40% by refactoring some
  extra big modules into more bite-sized pieces. #1806 (1.9.2)
• testtex:
  • Make the "thread workout" cases all honor --handle. #1778 (1.9.0)
  • Only prints detailed stats if -v is used, and new option
    --invalidate will invalidate the cache when starting each
    threadtimes trial. #1828 (1.9.2)
  • New --anisoratio lets you choose anisotropic shape for thread
    working tests, and make thread_workout samples twice as big to be more
    typical by interpolating mip levels. #1840 (1.9.2)
  • TextureSystem stats are printed as well as ImageCache. #1840 (1.9.2)
• iv no longer requires GLEW, using QOpenGLFunctions instead. #1840 (1.9.2)
• DICOM: Fix dcmtk build errors on some platforms. Also, the minimum dcmtk
  version we suport is 3.6.1. #1843 (1.9.2/1.8.8)
• Build fixes for Hurd OS. #1850 (1.9.2/1.8.8)
• Clean up leak sanitizer errors. #1855 (1.9.2)
• On Unix/Linux, add explicit DL library dependency to libOpenImageIO.so
  itself instead of only to the binaries and test utilities.
  #1860 (1.9.2/1.8.8)
• The build now bundles a sample OCIO config in testsuite/common so that we
  can do OCIO-based unit tests. #1870 (1.9.2)
• Properly find newer openjpeg 2.3. #1871 (1.9.2)
• Fix testsuite to be Python 2/3 agnostic. #1891 (1.9.2)
• Removed USE_PYTHON3 build flag, which didn't do anything. #1891 (1.9.2)
• The PYTHON_VERSION build variable is now better at selecting among
  several installed versions of Python, and all the tests should work fine
  with Python 3.x now. #2015 (1.9.4)
• Remove some lingering support for MSVS < 2013 (which we haven't advertised
  as working anyway). #1887 (1.9.2)
• Windows/MSVC build fix: use the /bigobj option on some large modules
  that need it. #1900, #1902 (1.8.10/1.9.2)
• Add up-to-date Nuke versions to FindNuke.cmake. #1920 (1.8.11, 1.9.2)
• Allow building against ffmpeg 4.0. #1926,#1936 (1.8.11, 1.9.2)
• Disable SSE for 32 bit Windows -- problematic build issues.
  #1933 (1.9.3, 1.8.12, 1.7.19)
• Fixes to the EMBEDPLUGINS=0 build case, which had at some point stopped
  working properly. #1942 (1.9.3)
• Improvements in finding the location of OpenJPEG with Macports.
  #1948 (1.8.12, 1.9.4)
• Improvement finding libraw properly on Windows. #1959 (1.9.4)
• Fix warnings to allow clean gcc8 builds. #1974 (1.9.4)
• Make sure we build properly for C++17. (1.9.4)
• Check properly for minimal FFMpeg version (2.6). #1981 (1.9.4)
• New build option GLIBCXX_USE_CXX11_ABI, when set to 0 will force the old
  gcc string ABI (even gcc 7+ where the new ABI is the default), and if set
  to 1 will force the new gcc string ABI (on gcc 5-6, where old ABI is the
  default). If not set at all, it will respect the default choice for that
  compiler. #1980 (1.9.4)
• TravisCI builds now use an abbreviated test matrix for most ordinary
  pushes of working branches, but the full test matrix for PRs or pushes
  to "master" or "RB" branches. #1983 (1.9.4)
• Support compilation by clang 7.0. #1995 (1.9.4)
• Support for building against OpenEXR 2.3. #2007 (1.9.4)
• Use OpenEXR pkgconfig if available. #2008 (1.9.4)
• Allow builds outside the source tree to pass testsuite. Defaults to
  finding test image directories such as oiio-images, openexr-images, and
  libtiffpic in the usual ".." from the main OIIO source directory, but now
  it can be overridden with the CMake variable OIIO_TESTSUITE_IMAGEDIR.
  #2026 (1.9.4)
• Remove stale python examples from src/python. They were untested,
  undocumented, and probably no longer worked against the current APIs.
  #2036 (1.9.4)
• Fixes for Windows when making Unicode builds, and fix Plugin::dlopen
  on Windows to properly support UTF-8 filenames. #1454 (2.0.1)
• Support added for OpenCV 4.0. (2.0.1)

Developer goodies / internals:

• Formatting with clang-format: All submissions are expected to be
  formatted using our standard clang-format rules. Please run
  make clang-format prior to submitting code. The TravisCI tests include
  one entry just to check that the formatting conforms, and will fail if it
  doesn't, printing the diffs that would bring it to proper formatting.
  (Note: for small changes, if you don't have clang-format locally, it's ok
  to submit, then use the diffs from the failures to fix it by hand and
  resubmit and update.) #2059,2064,2065,2067,2069.
• argparse.h:
  • Add pre- and post-option help printing callbacks. #1811 (1.9.2)
  • Changed to PIMPL to hide implementation from the public headers.
    Also modernized internals, no raw new/delete. #1858 (1.9.2)
• array_view.h:
  • Added begin(), end(), cbegin(), cend() methods, and new
    constructors from pointer pairs and from std::array. (1.9.0/1.8.6)
  • Deprecated, moved contents to span.h. You should change array_view<T>
    to span<T> and array_view<const T> to cspan<T>. #1956 (1.9.4)
• color.h: add guards to make this header safe for Cuda compilation.
  #1905 (1.9.2/1.8.10)
• filesystem.h:
  • IOProxy classes that can abstract file operations for custom I/O
    substitutions. #1931 (1.9.3)
  • Proper UTF-8 filenames for unique_path() and temp_directory(), and
    general UTF-8 cleanup/simplification. #1940 (1.9.3, 1.8.12, 1.7.19)
• fmath.h:
  • Now defines preprocessor symbol OIIO_FMATH_H so other files can
    easily detect if it has been included. (1.9.0/1.8.6)
  • Modify to allow Cuda compilation/use of this header. #1888,#1896
    (1.9.2/1.8.10)
  • Improve numeric approximation of fast_atan() and fast_atan2().
    #1943 (1.9.3)
  • fast_cbrt() is a fast approximate cube root (maximum error 8e-14,
    about 3 times faster than pow computes cube roots). #1955 (1.9.4)
• function_view.h: Overhauled fixed with an alternate implementation
  borrowed from LLVM. (1.9.4)
• hash.h: add guards to make this header safe for Cuda compilation.
  #1905 (1.9.2/1.8.10)
• imageio.h: convert_image() and parallel_convert_image have been
  simplified to remove optional alpha_channel and z_channel parameters
  that were never actually used. The old versions are still present but
  are deprecated. #2088 (2.0.1)
• parallel.h:
  • parallel_options passed to many functions. #1807 (1.9.2)
  • More careful avoidance of threads not recursively using the thread
    pool (which could lead to deadlocks). #1807 (1.9.2)
  • Internals refactor of task_set #1883 (1.9.2).
  • Make the thread pool better behaved in times if pool congestion -- if
    there are already way too many items in the task queue, the caller may
    do the work itself rather than add to the end and have to wait too
    long to get results. #1884 (1.9.2)
• paramlist.h:
  • ParamValue class has added get_int_indexed() and get_float_indexed()
    methods. #1773 (1.9.0/1.8.6)
  • ParamValue restructured to allow additional common data types to store
    internally rather than requre an allocation. #1812 (1.9.2)
  • New ParamList convenience methods: remove(), constains(),
    add_or_replace(). #1813 (1.9.2)
• platform.h:
  • New OIIO_FALLTHROUGH and OIIO_NODISCARD macros, and renamed
    OIIO_UNUSED_OK to OIIO_MAYBE_UNUSED (to match C++17 naming). #2041
• simd.h:
  • Fixed build break when AVX512VL is enabled. #1781 (1.9.0/1.8.6)
  • Minor fixes especially for avx512. #1846 (1.9.2/1.8.8) #1873,#1893
    (1.9.2)
• span.h:
  • Used to be array_view. Now it's span<> and span_strided. Also,
    cspan<T> is a handy alias for span<const T>. #1956 (1.9.4)
  • Added begin(), end(), cbegin(), cend() methods, and new
    constructors from pointer pairs and from std::array. (1.9.0/1.8.6)
  • Added == and != to span and span_strided. #2037 (1.9.4)
• strutil.h:
  • All string->numeric parsing and numeric->string formatting is now
    locale-independent and always uses '.' as decimal marker. #1796 (1.9.0)
  • New Strutil::stof(), stoi(), stoui(), stod() functions for
    easy parsing of strings to numbers. Also tests Strutil::string_is_int()
    and string_is_float(). #1796 (1.9.0)
  • New to_string<> utility template. #1814 (1.9.2)
  • Fix to strtof, strtod for non-C locales. #1918 (1.8.11, 1.9.2)
  • New iless() is case-insensitive locale-independent string_view
    ordering comparison. Also added StringIEqual, StringLess, StringILess
    functors. (1.9.4)
  • join() is now a template that can act on any iterable container of
    objects that allow stream output. #2033 (1.9.4)
  • New splits()/splitsv() that direction returns a vector of
    std::string or string_view, respectively. #2033 (1.9.4)
  • A second version of extract_from_list_string that directly returns
    a std::vector (instead of being passed as a param). #2033 (1.9.4)
  • parse_string now accepts single quotes as well as double quotes
    to enclose a quoted string. #2066 (2.0beta2)
  • Fix Strutil::vsnprintf detection of encoding errors on Windows. #2082
    (1.8.17/2.0.1)
• thread.h:
  • Reimplementaiton of spin_rw_mutex has much better performance when
    many threads are accessing at once, especially if most of them are
    reader threads. #1787 (1.9.0)
  • task_set: add wait_for_task() method that waits for just one task in
    the set to finish (versus wait() that waits for all). #1847 (1.9.2)
  • Fix rare crash in thread pool when lowering the number of threads.
    #2013 (1.9.4/1.8.15)
• unittest.h:
  • Made references to Strutil fully qualified in OIIO namespace, so that
    unittest.h can be more easily used outside of the OIIO codebase.
    #1791 (1.9.0)
  • OIIO_CHECK_EQUAL_APPROX - fix namespace ambiguity. #1998 (1.9.4)
  • OIIO_CHECK_EQUAL now can compare two std::vectors. #2033 (1.9.4)
  • Make unit test errors respect whether stdout is a terminal when
    deciding whether to print in color. #2045
• Extensive use of C++11 final and override decorators of virtual
  methods in our internals, especially ImageInput and ImageOutput.
  #1904 (1.9.2)

Notable documentation changes:

• A new LICENSE-THIRD-PARTY.md file reproduces the full open source licenses
  of all code that we distribute with OIIO, incorporate, or derive code from.
  They are have very similar license terms to the main OIIO license
  ("New BSD") -- MIT, Mozilla, Apache 2.0, or public domain. OIIO does not
  use any GPL or other "viral" licenesed code that would change license
  terms of any code that didn't come directly from those packages.
• The CHANGES.md file was getting truly enormous, so we have split the
  release notes from the 0.x and 1.x releases into separate files found
  in src/doc. So CHANGES.md only documents 2.0 and beyond.