OpenImageIO

Release 2.3.17 (1 Jul 2022) -- compared to 2.3.16

• TIFF: fix read problems with TIFF files with non-zero y offset. #3419
• Targa: Fix reading of tiny 1x1 2-bpp Targa 1.0 images. #3433 (2.3.17/2.2.21)
  (2.3.17/2.2.21)
• BMP: better detection of corrupted files with nonsensical image dimensions
  or total size. #3434 (2.3.17/2.2.21)
• BMP: protect against corrupted files that have palette indices out of bound.
  #3435 (2.3.17/2.2.21)
• TextureSystem: Fix typo that prevented "max_tile_channels" attribute from
  being set or retrieved. (2.3.17)
• ustring.h: ustring has added a from_hash() static method #3397, and a
  ustringhash helper class #3436. (2.3.17/2.2.21)
• benchmark.h: Alter the declaration of DoNotOptimize() so that it doesn't
  have compilation problems on some platforms. #3444 (2.3.17)
• Fix crash when ioproxy is passed to an image writer that doesn't support it.
  #3453 (2.3.17)
• Fix the "Makefile" wrapper to correctly recognize arm64 ("Apple silicon").
  #3456 (2.3.17)

Note: This is a patch release of an obsolete release family. The currently family
of releases enjoying full support is 2.3.

Release 2.2.21 (1 Jul 2022) -- compared to 2.2.20

• BMP: gain the ability to read some very old varieties of BMP files. #3375
• BMP: better detection of corrupted files with nonsensical image dimensions
  or total size. #3434
• BMP: protect against corrupted files that have palette indices out of bound.
  #3435
• ffmpeg: Support for ffmpeg 5.0. #3282
• ffmpeg: protect against possible double-free. #3376
• ffmpeg: make the supported file extension check be case-insensitive. This
  prevents movie files from being incorrectly unable to recognize their format
  if they have the wrong capitalization of the file extension. #3364
• hdr/rgbe files: Avoid possible Windows crash when dealing with characters
  with the high bit set. #3310
• TIFF: fix read problems with TIFF files with non-zero y offset. #3419
• Dev goodies: ustring has added a from_hash() static method #3397, and a
  ustringhash helper class #3436.
• simd.h fixes for armv7 and aarch32. #3361

Release 2.3.16 (1 Jun 2022) -- compared to 2.3.15

• UDIM texture lookups have been sped up by 5-8%. #3417
• Support for Intel llvm-based compiler 2022.1.0. #3407
• Internals: custom fmt formatters for vector types. #3367
• Fix compiler breaks when using some changes in fmtlib master (not yet
  released). #3416

Release 2.3.15 (1 May 2022) -- compared to 2.3.14

• JPEG: Better handling of PixelAspectRatio. #3366
• OpenEXR: Fix DWAA compression default level. #3387
• Perf: Huge speed-up of case-insensitive string comparisons (Strutil iequals,
  iless, starts_with, istarts_with, ends_with, iends_with), which also speeds
  up searches for attributes by name in ImageSpec and ParamValueList. #3388
• New ImageBufAlgo::st_warp() (and oiiotool --st_warp) perform warping of
  an image where a second image gives the (s,t) coordinates to look up from at
  every pixel. #3379
• Python: Add ImageSpec and ParamValueList method get_bytes_attribute(),
  which is like get_string_attribute(), but returns the string as a Python
  bytes object. In Python 3, strings are UTF-8, so this can be useful if you
  know that a string attribute might contain non-UTF8 data. #3396

Release 2.3.14 (1 Apr 2022) -- compared to 2.3.13

• Add support for UDIM pattern <uvtile> (used by Clarisse & V-Ray). #3358
• BMP: Support for additional (not exactly fully documented) varieties used by
  some Adobe apps. #3375
• Python: support uint8 array attributes in and out. This enables the proper
  Python access to "ICCProfile" metadata. #3378
• Improved precision in IBA::computePixelStats(). #3353
• ffmpeg reader not uses case-insensitive tests on file extensions. #3364
• Fix writing deep exrs when buffer datatype doesn't match the file. #3369
• Fix conflict between RESTful and Windows long path notations. #3372
• ffmpeg reader: take care against possible double-free of allocated memory
  crash upon destruction. #3376
• simd.h fixes for armv7 and aarch32. #3361
• Fix compiler warnings related to discrepancies between template declaration
  and redeclaration in simd.h and benchmark.h. #3350
• Suppress MacOS warnings about OpenGL depreation. #3380
• Now doing CI builds for Intel icc and icx compilers. #3355 #3363
• CI: Overhaul of yml file to be more clear and compact by using GHA
  "strategy" feature. #3356 #3365

Release 2.3.13 (1 Mar 2022) -- compared to 2.3.12

• Filesystem::searchpath_split better handling of empty paths. #3306
• New Strutil::isspace() is an isspace replacement that is safe for char
  values that are < 0. #3310
• Expose the Strutil::utf8_to_utf16() and utf16_to_utf8() utilities on
  non-Windows platforms (and also modernize their internals). #3307
• For the most important ImageInput, ImageOutput, and ImageBuf methods that
  take a filename, add new flavors that can accept a wstring as the
  filename. #3312 #3318
• PPM: properly report color space as Rec709 (as dictated by PPM spec). #3321
• PNG: more robust reporting of color space as sRGB in the absence of header
  fields contradicting this. #3321
• strutil.h: Split the including of fmt.h and some related logic into a
  separate detail/fmt.h. This is still included by strutil.h, so users
  should not notice any change. #3327
• Targa: Fix parsing of TGA 2.0 extension area. #3323
• Support building against FFmpeg 5.0. #3282
• oiiotool --pixelaspect : fix setting of "PixelAspectRatio", "XResolution",
  and "YResolution" attributes in the output file (were not set properly
  before). #3340

Release 2.3.12 (1 Feb 2022) -- compared to 2.3.11

• oiiotool: Don't give spurious warnings about no output when the --colorcount
  or --rangecheck commands are used. #3262
• oiiotool --pattern checker fixed behavior so that if only the checker
  width was specified but not the height, height will be equal to width. #3255
• oiiotool --point lets you set individual pixels. #3256
• Python: A new ImageBuf constructor was added that takes only a NumPy
  ndarray, and deduces the resolution, channels, and format (pixel data type)
  from the shape of the array. #3246
• Python: Implement ROI.copy(). #3253
• Python: ImageSpec and ParamValueList now support 'key' in spec,
  del spec['key'], and spec.get('key', defaultval) to more fully emulate
  Python dict syntax for manipulating metadata. #3252 (2.3.12/2.4.0)
• Python bug fix: fix clamp() when the min or max are just a float. Now
  it uses that one value for all channels, instead of using it only for
  the first channel. #3265
• ImageSpec gained an additional constructor that takes a string
  representation of the pixel data type (where it used to insist on a
  TypeDesc), such as ImageSpec(640, 480, 3, "uint8"). This is especially
  helpful for new/casual users, or when you want code to be maximally
  readable. #3245
• Imagepec::getattribute() new query token "format" can retrieve the pixel
  data type. #3247
• IBA::make_texture(): ensure that "maketx:ignore_unassoc" is honored. #3269
• Support an additional UDIM pattern <UVTILE>, which is specified by
  MaterialX. #3280
• TIFF: support 16-bit palette images. #3260
• TIFF: Gracefully handle missing ExtraSamples tag. #3287
• Targa: Better interpretation of TGA 1.0 files with alpha that is zero
  everywhere. Be more consistent with Targa attributes all being called
  "targa:foo". Add "targa:version" to reveal whether the file was TGA 1.0
  or 2.0 version of the format. #3279
• simd.h: Better guards to make it safe to include from Cuda. #3291 #3292
• Fix bugs in the build_opencolorio.bash script, did not correctly handle
  installation into custom directories. #3278
• Fixes to FindOpenColorIO.cmake module, now it prefers an OCIO exported cmake
  config (for OCIO 2.1+) unless OPENCOLORIO_NO_CONFIG=ON is set. #3278
• Docs: The ImageBufAlgo chapter now has examples for C++, Python, and
  oiiotool for almost every operation. #3263

Release 2.2.20 (1 Feb 2022) -- compared to 2.2.19

• Fix some address sanitizer failures. #3160
• Build/CI: Deal with OpenColor renaming its master branch to main. #3169
• Windows: Fix error when compiling with MSVC. #3168
• Fix excessive memory usage when saving EXR with many channels. #3176
• TIFF: now works for 16-bit palette images. #3260
• Fix ImageBuf::read bug for images of mixed per-channel data types. #3088

Release 2.3.11 (1 Jan 2022) -- compared to 2.3.10

• JPEG2000: enable multithreading for decoding (if using OpenJPEG >= 2.2)
  and encoding (if using OpenJPEG >= 2.4). This speeds up JPEG-2000 I/O
  by 3-4x. #2225
• TIFF: automatically switch to "bigtiff" format for >4GB images. #3158
• Security: New global OIIO attributes "limits:channels" (default: 1024) and
  "limits:imagesize_MB" (default: 32768, or 32 GB) are intended to reject
  input files that exceed these limits, on the assumption that they are either
  corrupt or maliciously constructed, and would, if read, lead to absurd
  allocations, crashes, or other mayhem. Apps may lower or raise these limits
  if they know that a legitimate input image exceeds these limits. Currently,
  only the TIFF reader checks these limits, but others will be modified to
  honor the limits over time. #3230
• TextureSystem: enhance safety/correctness for untiled images that exceed
  2GB size (there was an integer overflow problem in computing offsets within
  tiles). #3232
• Cleanup: Get rid of an obsolete header c-imageio.h that was experimental
  and the functions declared therein were not implemented in this release.
  #3237
• Build: rely on env variable "OpenImageIO_CI" to tell us if we are running in
  a CI environment, not the more generic "CI" which could be set for other
  reasons in some environments. #3211
• Build: Improvements in how we find OpenVDB. #3216
• Build: If CMake variable BUILD_TESTING is OFF, don't do any automatic
  downloading of missing test data. #3227
• Dev: Add Strutil::edit_distance(). #3229
• Docs: Clean up intra-document section references, new explanations about
  input and output configuration hints, more code examples in both C++ and
  Python (especially for the ImageInput and ImageOutput chapters). #3238 #3244

Release 2.3.10.1 (7 Dec 2021) -- compared to 2.3.10.0

• Build: restore code that finds Jasper when using statically-linked libraw.
  #3210
• Build/test: Gracefully handle failing to find git for test data download.
  #3212
• fmath.h: bit_cast specialization should take refs, like the template.
  This made warnings for some compilrs. #3213
• Build: make sure to properly use the tbb target if it exists. #3214

Release 2.3.10 (1 Dec 2021) -- compared to 2.3.9

New (non-compatibility-breaking) features:

• TextureSystem: add feature for stochastic mipmap interpolation. This adds
  new interpolation modes "StochasticTrilinear" and "StochasticAniso", which
  in conjunction with the "rnd" field in TextureOpt, will stochastically
  choose between bracketing MIPmap levels rather than interpolating them. This
  reduces texture lookup cost by up to 40%, but it's only useful in the
  context of a renderer that uses many samples per pixel. #3127
• maketx/make_texture() now supports options to store Gaussian forward and
  inverse transform lookup tables in image metadata (must be OpenEXR textures
  for this to work) to aid writing shaders that use histogram-preserving
  blending of texture tiling. This is controlled by new maketx arguments
  --cdf, --cdfsigma, --sdfbits, or for IBA::make_texture() by using
  hints maketx:cdf, maketx:cdfsigma, and maketx:cdfbits. #3159
• oiitool --oiioattrib can set "global" OIIO control attributes for
  an oiiotool run (equivalent of calling OIIO::attribute()). #3171
• oiiotool --repremult exposes the previously existing IBA::repremult().
  The guidance here is that --premult should be used for one-time conversion
  of "unassociated alpha/unpremultiplied color" to associated/premultiplied,
  but when you are starting with a premultiplied image and have a sequence of
  unpremultiply, doing some adjustment in unpremultiplied space, then
  re-premultiplying, it's --repremult you want as the last step, because it
  preserves alpha = 0, color > 0 data without crushing it to black. #3192
• oiiotool --saturate and IBA::saturate() can adjust saturation level of a
  color image. #3190
• When building against OpenEXR >= 3.1.3, our OpenEXR output now supports
  specifying the zip compression level (for example, by passing the
  "compression" metadata as "zip:4"). Also note than when using OpenEXR >=
  3.1.3, the default zip compression has been changed from 6 to 4, which
  writes compressed files significantly (tens of percent) faster, but only
  increases compressed file size by 1-2%. #3157
• Improved image dithering facilities: When dithering is chosen, it now
  happens any time you reduce >8 bits to <= 8 bits (not just when converting
  from float or half); change the dither pattern from hashed to blue noise,
  which looks MUCH better (beware slightly changed appearance); IBA::noise()
  and oiiotool --noise now take "blue" as a noise name, giving a blue noise
  pattern; IBA::bluenoise_image() returns a reference to a stored periodic
  blue noise iamge; oiiotool -d now lets you ask for "uint6", "uint4",
  "uint2", and "uint1" bit depths, for formats that support them. #3141
• New global OIIO attribute "try_all_readers" can be set to 0 if you want to
  override the default behavior and specifically NOT try any format readers
  that don't match the file extension of an input image (usually, it will try
  that one first, but it if fails to open the file, all known file readers
  will be tried in case the file is just misnamed, but sometimes you don't
  want it to do that). #3172
• Raise the default ImageCache default tile cache from 256MB to 1GB. This
  should improve performance for some operations involving large images or
  images with many channels. #3180
• IOProxy support has been added to JPEG output (already was supported for
  JPEG input) and for GIF input and output. #3181 #3182
• oiiotool --maxchan and --minchan, and IBA::maxchan() and minchan()
  turn an N-channel image into a 1-channel images that for each pixel,
  contains the maximum value in any channel of the original for that pixel.
  #3198

Bug fixes:

• Fix oiiotool --invert to avoid losing the alpha channel values. #3191
• Fix excessive memory usage when saving EXR files with many channels. #3176
• WebP: Fix previous failure to properly set the "oiio:LoopCount" metadata
  for animated webp images. #3183
• Targa: Better detection/safety when reading corrupted files, and fixed
  bug when reading x-flipped images. #3162
• RLA: better guards against malformed input. #3163
• Fix oiiotool bug when autocropping output images when the entire pixel data
  window is in the negative coordinate region. #3164
• oiiotool improved detection of file reading failures. #3165
• DDS: Don't set "texturetype" metadata, it should always have been only
  "textureformat". Also, add unit testing of DDS to the testsuite. #3200
• When textures are created with the "monochrome-detect" feature enabled,
  which turns RGB textures where all channels are always equal into true
  single channel greyscale, the single channel that results is now correctly
  named "Y" instead of leaving it as "R" (it's not red, it's luminance).
  #3205

Build fixes and developer goodies:

• Update internal stb_printf implementation (avoids some sanitizer alerts).
  #3160
• Fixes for MSVS compile. #3168
• Add Cuda host/device decorations to TypeDesc methods to make them GPU
  friendly. #3188
• TypeDesc: The constructor from a string now accepts "box2f" and "box3f"
  as synonyms for "box2" and "box3", respectively. #3183
• New Strutil::parse_values, scan_values, scan_datetime, parse_line. #3173
  #3177
• New Filesystem::write_binary_file() utility function. #3199
• New build option -DTIME_COMMANDS=ON will print time to compile each module
  (for investigating build performance; only useful when building with
  CMAKE_BUILD_PARALLEL_LEVEL=1). #3194

Release 2.3.9.1 (1 Nov 2021) -- compared to 2.3.8

• OpenEXR: When building against OpenEXR 3.1+ and when the global OIIO
  attribute "openexr:core" is set to nonzero, do more efficient multithreaded
  reading of OpenEXR files. #3107
• oiiotool --dumpdata:C=name causes the dumped image data to be formatted
  with the syntax of a C array. #3136
• oiiotool: Allow quotes in command modifiers. #3112
• jpeg input: remove stray debugging output to console. #3134
• HEIF: Handle images with unassociated alpha. #3146
• Targa: improved error detection for read errors and corrupted files. #3120
• raw: When using libraw 0.21+, now support new color space names "DCE-P3",
  "Rec2020", and "sRGB-linear", and "ProPhoto-linear". Fix incorrect gamma
  values for "ProPhoto". #3123 #3153 Fixes to work with the libraw 202110
  snapshot. #3143
• TIFF: honor zip compression quality request when writing TIFF. #3110
• Fix broken oiiotool --dumpdata output. #3131
• Fix possible bad data alignment and SIMD assumptions inside TextureSystems
  internals. #3145
• Field3D, which is no longer actively supported, now has support disabled in
  the default build. To enable Field3D support, you must build with
  -DENABLE_FIELD3D=1. Note that we expect it to be no longer supported at
  all, beginning with OIIO 2.4. #3140
• Build: Address new warnings revealed by clang 13. #3122
• Build: Fix when building with Clang on big-endian architectures. #3133
• Build: Fix occasional build breaks related to OpenCV headers. #3135
• Build: Improvements to NetBSD and OpenBSD support. #3137.
• docs: Add an oiiotool example of putting a border around an image. #3138
• docs: Fix explanation of ImageCache "falure_retries" attribute. #3147

Release 2.2.19 (1 Nov 2021) -- compared to 2.2.18

• Better catching of exceptions thrown by OCIO 1.x if it encounters 2.0 config
  files. #3089
• Address new warnings revealed by clang 13. #3122
• Fixed some minor python binding bugs. #3074 #3094
• Fix when building with Clang on big-endian architectures. #3133
• Fix occasional build breaks related to OpenCV headers. #3135
• Improvements to NetBSD and OpenBSD support. #3137.
• Fixes to work with the libraw 202110 snapshot. #3143

Note: This is an update to the obsolete 2.2 release family. The current
fully-supported release family is 2.3.

Release 2.3.8 (1 Oct 2021) -- compared to 2.3.7

• Fix ImageBuf::read() bug for images of mixed per-channel data types. #3088
• Fix crash that could happen with invalidly numbered UDIM files. #3116
• Better catching of exceptions and other error handling when OpenColorIO 1.x
  is used but encounters an OpenColorIO 2.x config file. #3089 #3092 #3095
• Ensure that OpenColorIO doesn't send info messages to the console if no
  config is found. #3113
• Fix: make sure ImageSpec deserialization works for arbitrary attribs. #3066
• oiiotool -ch now has greatly reduced cost (no useless allocations or
  copies) when the channel order and names don't change. #3068
• oiiotool --runstats is now much better about correctly attributing I/O
  time to -i instead of to the subsequent operations that triggers the
  read. #3073
• TIFF output now supports IO proxies. #3075 #3077
• Improved finding of fonts (by IBA::render_text and oiiotool --text). It now
  honors environment variable $OPENIMAGEIO_FONTS and global OIIO attribute
  "font_searchpath" to list directories to be searched when fonts are needed.
  #3096
• We no longer install a FindOpenImageIO.cmake module. It was incomplete, out
  of date, and wholly unnecessary now that we correctly export a config file
  OpenImageIOConfig.cmake and friends. #3098
• When building against OpenEXR 3.1+, use of the OpenEXRCore library no longer
  requires a build-time option, but instead is always available (though off by
  default) and can be enabled by an application setting the OIIO global
  attribute "openexr:core" to 1. #3100
• dev: Timer::add_seconds and add_ticks methods. #3070
• dev: Add round_down_to_multiple() and improve round_to_multiple() to
  correctly handle cases where the value is less than 0. #3104

Release 2.3 (1 Sept 2021) -- compared to 2.2

New minimum dependencies and compatibility changes:

• C++ standard: C++14 is now the minimum (gcc 6.1 - 11.2, clang 3.4 - 12,
  MSVS 2017 - 2019, icc 17+). The default C++ standard mode, if none is
  explicitly specified, is now C++14. #2918 (2.3.4) #2955 #2977 (2.3.5)
• FFMmpeg (optional) dependency minimum is now 3.0 (raised from 2.6). #2994
• OpenCV (optional) dependency minimum is now 3.0 (raised from 2.0). #3015

New major features and public API changes:

• Changes and clarifications to geterror() in all classes that have one:
  • geterror() now takes an optional clear parameter controls if the
    pending error is cleared as well as retrieved (default true).
  • All classes with geterror() are ensured to have has_error().
  • geterror() appends to the pending error message, if one is already
    present but not retrieved. Appending errors will be automatically
    separated by newlines, also any newline at the end of a geterror
    call will be stripped. #2740 (2.3.0.1)
• ImageInput error messages are now thread-specific. This prevents multiple
  threads simultaneously reading from the same ImageInput from getting the
  other thread's error messages. But it means you must always retrieve an
  error with getmessage() from the same thread that made the read call
  that generated the error. #2752 (2.3.0.1)
• ImageInput and ImageOutput have removed several fields (like the m_mutex)
  and instead now use a PIMPL idiom that hides internals better from the
  ABI. If you had a custom ImageInput/Output class that directly accessed
  the m_mutex, replace it instead with calls to ImageInput::lock() and
  unlock() (and ImageOutput). #2752 (2.3.1.0)
• Clarify that ImageBuf methods subimage(), nsubimages(), miplevel(),
  nmipevels(), and file_format_name() refer to the file that an ImageBuf
  was read from, and are thus only meaningful for ImageBuf's that directly
  read from files. "Computed" or copied ImageBuf's are solo images in memory
  that no longer correspond to a file even if the input to the computation
  was from a file. #2759 (2.3.0.1/2.2.9)
• New get_extension_map() returns a map that list all image file formats
  and their presumed file extensions. #2904 (2.3.4)
• Overhaul of how we handle "thumbnails" (small preview images embedded in
  the headers of some image file formats) #3021 (2.3.7):
  • Thumbnails no longer embedded in the ImageSpec as "thumbnail_image"
    metadata of a big character array blob.
  • ImageInput, ImageBuf, and ImageCache have varieties of get_thumbnail()
    method that returns the thumbnail as an ImageBuf. If this is not
    called, ideally the reader will not undergo the expense of reading and
    storing thumbnail data.
  • ImageOutput and ImageBuf have set_thumbnail() to communicate a
    thumbnail image as an ImageBuf.
  • The net effect is that reading files and retrieving ImageSpec's should
    be more efficient for apps that don't have a use for thumbnails.
• Significant improvements to UDIM handling in ImageCache/TextureSystem:
  • New UDIM texture name patterns recognized: %(UDIM)d is the Houdini
    convention, and _u##v## is for Animal Logic's internal renderer.
    #3006 (2.2.16/2.3.6)
  • It is now permissible for get_texture_info()/get_image_info() to
    retrieve metadata for UDIM patterns -- previously it failed, but now
    it will succeed in retrieving any metadata as long as it is present
    and has the same value in all the matching "UDIM tile" panels. #3049
    (2.3.7)
  • TextureSystem now exposes methods is_udim(), resolve_udim(),
    inventory_udim() that should be helpful for apps dealing with UDIM
    textures. See the documentation for details. #3054 (2.3.7)
  • Performance improvements when using UDIM textures. #3049 (2.3.7)
• ImageBuf, when "wrapping" an app-owned buffer, now allows explicit
  specification of stride lengths. This allows you to wrap a buffer that
  has internal padding or other non-contiguous spacing of pixels, scanlines,
  or image planes. #3022 (2.3.6)
• oiiotool new commands and options:
  • --pastemeta takes two images as arguments, and appends all the
    metadata (only) from the first image onto the second image's pixels
    and metadata, producing a combined image. #2708 (2.3.0.0)
  • --chappend and --siappend both allow an optional modifier :n=
    to specify the number of images from the stack to be combined
    (default n=2). #2709 (2.3.0.0)
  • --fit now takes an additional optional modifier: fillmode= with
    choices of letterbox (default), width, and height. #2784 (2.3.2.0)
  • --autocc now has an optional modifier :unpremult=1 that causes
    any color transformations related to autocc to unpremultiply by alpha
    before the transformation and then re-premultiply afterwards, exactly
    the same control that exists for individual --colorconvert but never
    existed for autocc. #2890 (2.3.3)
  • --list-formats lists all the image file formats that OIIO knows
    about, and for each, the associated file extensions. #2904 (2.3.4)
  • --skip-bad-frames causes an error (such as an input file not being
    found) to simply skip to the next frame in the frame range loop, rather
    that immediately exiting. #2905 (2.3.4)
  • When doing expression substitution, {getattribute(name)} will be
    replaced by the value that OIIO::getattribute(name, ...) would
    retrieve. #2932 (2.3.4)
  • --missingfile (which takes a subsequent argument of error, black,
    or checker) determines the behavior when an input image file is
    missing (the file does not exist at all, does not include cases where
    the file exists but there are read errors). The default value of
    error matches the old behavior: report an error and terminate all
    command processing. Other choices of black or checker will continue
    processing but substitute either a black or checkerboard image for the
    missing file. #2949 (2.3.4)
  • --printinfo prints verbose metadata info about the current top
    image on the stack. #3056 (2.3.7)
  • --printstats prints statistics about the current top image on the
    stack, and optionally can limit the statistics to a rectangular
    subregion of the image. #3056 (2.3.7)
  • Expanded expression now support the following new metadata notations:
    META, METABRIEF, and STATS. #3025 (2.3.6)
  • --mosaic now takes optional fit=WxH modifier that lets you set the
    "cell" size, with all constituent images resized as needed (similarly
    to if you used --fit on each individually). This is helpful for
    creating contact sheets without having to resize separately or to know
    the resolution of the original images. #3026 (2.3.6)
  • REMOVED --histogram which had been deprecated and undocumented since
    before OIIO 2.0. #3029 (2.3.6)
• Python bindings:
  • When transferring blocks of pixels (e.g., ImageInput.read_image()
    or ImageOutput.write_scanline()), "half" pixels ended up mangled
    into uint16, but now they use the correct numpy.float16 type. #2694
    (2.3.0.0)
  • The value passed to attribute(name, typedesc, value) can now be a
    tuple, list, numpy array, or scalar value. #2695 (2.3.0.0)
  • Added Python bindings for the TextureSystem. #2842
  • Add the previously (inadvertently) omitted enum value for
    ImageBufAlgo.MakeTxBumpWithSlopes. #2951 (2.3.4)
• Environment variables:
  • CUE_THREADS env variable, if set, is now honored to set the default
    size of the OIIO thread pool. This helps to make OIIO-based apps be
    good citizens when run as OpenCue jobs. #3038 (2.3.7)
• Library organization:
  • All the utility classes are now in libOpenImageIO_Util only and
    libOpenImageIO depends on and links to libOpenImageIO_Util, rather
    than the utility classes being defined separately in both libraries.
    #2906 (2.3.4)
  • Include file change: imagebuf.h no longer includes imagecache.h.
    It should never have needed to, since it didn't need any of its
    declarations. Other code that should always have included both headers,
    but inadvertently only included imagebuf.h, may now need to add an
    explicit #include <OpenImageIO/imagecache.h>. #3036 (2.3.7)
• Important change to Makefile wrapper: We have a 'Makefile' that just wraps
  cmake for ease of use. This has changed its practice of putting builds in
  build/ARCH to just build (and local installs from dist/ARCH to dist)
  to better match common cmake practice. So where your local builds end up may
  shift a bit. #3057 (2.3.7)
• Deprecations: A number of long-deprecated functions and methods have been
  given deprecation attributes that will cause warnings if you use them. They
  will eventually be removed entirely at the next "breaking" release (3.0,
  whenever that is). #3032 (2.3.7)

Performance improvements:

• Speed up BMP reading by eliminating wasteful allocation and copying
  done on each scanline read. #2934 (2.3.4)
• For IBA::to_OpenCV(), improve efficiency for certain image copies. #2944
  (2.3.4)
• Fix runaway parsing time for pathological XMP metadata. #2968 (2.3.5/2.2.15)
• TextureSystem reduction in overhead when using UDIM textures. #3049 (2.3.7)

Fixes and feature enhancements:

• Fix a situation where if an ImageBuf backed by an ImageCache reads an
  image, then the image changes on disk, then another ImageBuf or ImageCache
  tries to read it, it could end up with the old version. This involved some
  strategic cache invalidation when ImageBuf's write images to disk. #2696
  (2.3.0.0)
• Improve parsing of XMP records in metadata: more correct handling of
  lists/sequences, better inference of types that look like int or float
  (rather than forcing unknown fields into strings), fixed bugs in parsing
  rational values. #2865 (2.2.12/2.3.3)
• ImageBuf/ImageBufAlgo:
  • IBA::contrast_remap() fixes bug that could crash for very large
    images #2704 (2.3.0.0)
  • Fix stack overflow crash in IBA::colorconvert of unusually wide images.
    #2716 (2.3.0.1/2.2.8)
  • Fix IBA::make_texture() incorrectly setting tile sizes. #2737
    (2.3.0.1/2.2.8)
  • IBA::make_texture() now correctly propagates error messages, which
    can be retrieved via the global OIIO::geterror() call. #2747 (2.3.0.1)
  • IBA::fit() now takes a fillmode parameter that controls exactly
    how the resize will occur in cases where the aspect ratio of the new
    frame doesn't exactly match that of the source image. The old way (and
    the default value now) is "letterbox", but now also "width" and
    "height" modes are allowed. See fit() documentation for details.
    #2784 (2.3.2.0)
  • IBA::fit() fixes some slight ringing at image edges when using a
    combination of black wrap mode and a filter with negative lobes.
    #2787 (2.3.2.0)
  • Fix: ociolook() and ociofiletransform() internally reversed the
    order of their inverse and unpremult arguments, making it hard to
    select the inverse transformation. #2844 (2.3.3/2.2.11)
  • Fix crash for certain calls to ImageBuf::set_write_format() when
    writing files that support per-channel data types. #2885 (2.3.3)
  • Fix: IBA::fillholes_pushpull did not correctly understand which
    channel was alpha when generating subimages. #2939 (2.3.4)
  • Fix: IBA::render_text did not properly account for non-1 alpha of
    the text color. #2981 (2.3.5/2.2.15)
  • IBA::colorconvert, colormatrixtransform, ociolook, ociodisplay,
    and ociofiletransform have been fixed so that if input is more than
    4 channels, the additional channels will be copied unchanged, rather
    than inadvertently set to 0. #2987 (2.3.5/2.2.16)
  • Thread safety has been improved for ImageBuf::init_spec() and read()
    (though no problem was ever reported in the wild). #3018 (2.3.6)
  • render_text() now accepts text strings with embedded linefeeds and
    will turn them into multiple lines of rendered text. #3024 (2.3.6)
• ImageCache/TextureSystem/maketx:
  • Fix ImageCache bug: add_tile/get_tile not properly honoring when
    chend < chbegin it should get all channels. #2742 (2.3.0.1/2.2.8)
  • ImageBufAlgo::make_texture() (as well as maketx and oiiotool -otex)
    clamp half values to their maximum finite range to prevent very large
    float inputs turning into Inf when saved as half values. #2891 (2.3.3)
  • Fix crash when sampling non-zero-base channels. #2962 (2.3.5/2.2.15)
  • IBA::make_texture() rejects "deep" input images, since they cannot
    be made into textures. Doing this could previously crash. #2991
    (2.3.5/2.2.16)
  • maketx and oiiotool -otex have fixed double printing of error
    messages #2992 (2.3.5).
  • Making bumpslopes textures now allows scaled slopes UV normalization.
    This is exposed as new maketx argument --uvsopes_scale, or passing
    the attribute uvslopes_scale (int) to make_texture(). #3012 (2.3.6)
  • There is a new variety of ImageCache::invalidate() that takes an
    ImageHandle* argument instead of a filename. #3035 (2.3.7)
• oiiotool:
  • --resize of images with multi-subimages could crash. #2711 (2.3.0.0)
  • Improve oiiotool's guessing about the desired output format based on
    inputs (in the absence of -d to specify the format). #2717
    (2.3.0.1/2.2.8)
  • --text now accepts text strings with embedded linefeeds and
    will turn them into multiple lines of rendered text. #3024 (2.3.6)
• BMP
  • Fix reading BMP images with bottom-to-top row order. #2776
    (2.3.1.1/2.2.9)
  • Correctly read BMP images of the older V1 variety without crashing.
    #2899 (2.3.3/2.2.13)
  • Fix error in the rightmost pixel column when reading 4 bpp images with
    odd horizontal resolution. #2899 (2.3.3/2.2.13)
  • BMP reads now return metadata "bmp:bitsperpixel" (only when not 24
    or 32) to indicate non-whole-byte channels sizes in the file, and
    "bmp:version" to indicate the version of the BMP format. #2899
    (2.3.3/2.2.13)
  • Speed up BMP reading by eliminating wasteful allocation and copying
    done on each scanline read. #2934 (2.3.4)
  • For BMP files that are 8 bits per pixel and all palette entries have
    R == G == B values, read the file as an 8-bit grayscale image instead
    of needlessly promoting to a full RGB 3-channel image. #2943 (2.3.4)
  • Full support for reading RLE-compressed BMP images. #2976 (2.3.5/2.2.15)
  • Write single channel BMP as 8 bit palette images. #2976 (2.3.5/2.2.15)
• DPX
  • Output to DPX files now supports IOProxy. (Input already did.) #3013
    (2.2.17/2.3.6)
• FFMpeg/movies:
  • Avoid potential crash when a frame can't be read. #2693 (2.3.0.0)
  • Several varieties of encoded videos with 12 bits per channel and
    having alpha were incorrectly reported as 8 bits and without alpha.
    #2989 (2.3.5/2.2.16)
• GIF
  • Support UTF-8 filenames on Windows for GIF files. #2777 (2.3.2.0)
  • Fix error checking for non-existant GIF files. #2886 (2.3.3)
• HEIC
  • Fix error decoding Apple HEIF files. #2794/#2809 (2.3.2.0)
  • Better valid_file() check for HEIF. #2810 (2.3.2.0)
  • Enabled AVIF decoding of heic files (requires libheif >= 1.7 and for
    it to have been built with an AV1 encoder/decoder). #2811 #2812 #2814
    #2818 (2.3.3.0)
• IFF
  • Fix broken reads of 16 bit iff files. #2736 (2.3.0.1/2.2.8)
• JPEG
  • Allow reading of JPEG files with mildly corrupted headers. #2927 (2.3.4)
• OpenEXR:
  • Fix rare crash that was possible when multithreaded writing openexr
    files. #2781 (2.3.2.0)
  • Improved error reporting when errors are encountered writing OpenEXR
    files. #2783 (2.3.2.0)
  • Fix potential crash parsing OpenEXR header that contains Rational
    attributes with certain values. #2791 (2.2.10/2.3.2)
  • EXPERIMENTAL: When building against OpenEXR 3.1, the OIIO CMake option
    -DOIIO_USE_EXR_C_API=ON will use a new OpenEXR API that we think
    will allow higher performance texture reads in a multithreaded app.
    This has not yet been benchmarked or tested thoroughly. #3009 #3027
    (2.3.6)
  • The reader no longer renames file metadata "version" into the special
    "openexr:version" that would indicate the OpenEXR file format version
    of the file. That's never what it meant! It really was just arbitrary
    metadata. #3044 (2.3.7)
• PNG
  • Read Exif data from PNG files. #2767 (2.3.1.1/2.2.9)
• PSD
  • Add support for reading ISD profiles for PSD input. #2788 (2.3.2.0)
  • Fix loading PSB files with cinf tags. #2877 (2.2.12/2.3.2)
• RAW:
  • Additional input configuration hints to control options in the
    underlying LibRaw: "raw:user_flip" #2769 (2.3.1.0)
    "raw:balance_clamped", "raw:apply_scene_linear_scale",
    "raw:camera_to_scene_linear_scale" #3045 (2.3.7)
  • Correctly handle RAW files with Unicode filenames on Windows. #2888
    (2.3.3)
  • Fix garbled output when raw:Demosaic hint is "none". #3045 (2.3.7)
• Targa
  • Fix potential crash when reading files with no thumbnail. #2903
    (2.3.3/2.2.13)
  • Fix alpha handling for some files. #3019 (2.3.6)
• TIFF:
  • Fix broken reads of multi-subimage non-spectral files (such as
    photometric YCbCr mode). #2692 (2.3.0.0)
  • Fix spec() and spec_dimensions() for MIPmapped TIFF files, they did
    not recognize being asked to return the specs for invalid subimage
    indices. #2723 (2.3.0.1/2.2.7)
  • Add ability to output 1bpp TIFF files. #2722 (2.3.0.1/2.2.7)
  • Fix reading TIFF files with "separate" planarconfig and rowsperstrip
    more than 1. #2757 (2.3.0.1/2.2.9)
  • Fix incorrect reading of tiled TIFF files using certain rare TIFF
    features that also have a vertical resolution that is not a whole
    multiple of the tile size. #2895 (2.3.3/2.2.13)
  • Support IOProxy for reading TIFF files. #2921 (2.3.4)
  • TIFF plugin now properly honors caller request for single-threaded
    operation. #3016 (2.3.6)
• WebP:
  • Add support for requesting compression "lossless". #2726 (2.3.0.1/2.2.8)
  • Input improvements including: RGB images are kept as RGB instead of
    always adding alpha; more efficient by skipping alpha premultiplication
    when it's unnecessary; now can read animated WebP images as
    multi-subimage files. #2730 (2.3.0.1/2.2.8)
• Fix memory leak during decoding of some invalid Exif blocks. #2824
• Fix possible divide-by-zero error in read_image/read_scanlines for invalid
  image sizes (usually from corrupt files). #2983 (2.3.5/2.2.16)

Developer goodies / internals:

• Internals now use C++11 final keywords wherever applicable. #2734
  (2.3.0.1)
• More internals conversion of old Strutil::sprintf to Strutil::fmt::format
  and related changes. #2889 (2.3.3)
• Redundant format conversion code removed from imageio.cpp #2907 (2.3.4)
• Eliminate direct output to std::cerr and std::cout by library calls.
  #2923 (2.3.4)
• argparse.h:
  • ArgParse::abort() lets the response to a command line argument signal
    that no further arguments should be parsed. #2820 (2.3.3/2.2.11)
• color.h:
  • New ColorConfig::OpenColorIO_version_hex() returns the hex code for
    the version of OCIO we are using (0 for no OCIO support). #2849 (2.3.3)
• farmhash.h:
  • Clean up all non-namespaced preprocessor symbols that are set
    by this header and may pollute the caller's symbols. #3002 (2.2.16/2.3.6)
• filesystem.h:
  • New Filesystem::generic_filepath() returnss a filepath in generic
    format (not OS specific). #2819 (2.3.3/2.2.11)
  • Improve exception safety in Filesystem directory iteration. #2998
    (2.2.16/2.3.6)
  • New filename_to_regex() makes a filename "safe" to use as a
    regex pattern (for example, properly backslashing any . characters).
    #3046 (2.3.7)
  • Some methods that used to take a const std::string& now take a
    string_view. #3047 (2.3.7)
• fmath.h:
  • Use CPU intrinsics to speed up swap_ending (by 8-15x when swapping
    bytes of large arrays). #2763 (2.3.1.0)
• hash.h:
  • farmhash::inlined::Hash now is constexpr and works for Cuda.
    #2843 (2.3.3) #2914 #2930 (2.3.4)
• oiioversion.h:
  • New macros OIIO_VERSION_GREATER_EQUAL and OIIO_VERSION_LESS.
    #2831 (2.3.3/2.2.11)
• platform.h:
  • New macro OIIO_INLINE_CONSTEXPR, equivalent to inline constexpr for
    C++17, but just constexpr for C++ <= 14. #2832 (2.3.3/2.2.11)
• simd.h:
  • Fix incorrect ARM NEON code in simd.h. #2739 (2.3.0.1/2.2.8)
• span.h:
  • std::size() and std::ssize() should work with OIIO::span now.
    #2827 (2.3.3/2.2.11)
• string_view.h:
  • std::size() and std::ssize() should work with OIIO::string_view
    now. #2827 (2.3.3/2.2.11)
  • More thorough constexr of string_view methods. #2841 (2.3.3)
• strongparam.h:
  • New StrongParam helper for disambiguating parameters. #2735 (2.3.2)
• strutil.h:
  • Strutil splits() and splitsv() should return no pieces when passed
    an empty string. (It previously erroneously returned one piece
    consisting of an empty string.) #2712 (2.3.0.0)
  • Fix build break when strutil.h was included in Cuda 10.1 code. #2743
    (2.3.0.1/2.2.8)
  • strhash() is now constexpr for C++14 and higher. #2843 (2.3.3)
  • New Strutil functions: find, rfind, ifind, irfind #2960 (2.3.4/2.2.14)
    contains_any_char #3034 (2.3.7) rcontains, ircontains #3053 (2.3.7)
• sysutil.h:
  • Extend getenv() to take a default if the environment variable is not
    found. #3037 #3040 (2.3.7)
• typedesc.h:
  • TypeDesc::basetype_merge(a,b) returns a BASETYPE having the
    precision and range to hold the basetypes of either a or b.
    #2715 (2.3.0.0)
  • TypeDesc can now describe 2D and 3D bounding boxes, as arrays of 2
    VEC2 aggregates (for 2D) or VEC3 aggregates (for 3D) with "BOX"
    semantic. The shorthand for these are TypeBox2, TypeBox3 (for
    float), and TypeBox2i and TypeBox3i for integer or pixel coordinte
    boxes. #3008 (2.2.17/2.3.6)
• unordered_map_concurrent.h:
  • New methods find_or_insert, nobin_mask(). #2867 (2.2.12/2.3.3)
• ustring.h:
  • ustring internals now have a way to ask for the list of ustrings whose
    hashses collided. #2786 (2.2.11/2.3.2.0)
  • ustring now guarantees that no two ustrings will return the exact same
    value for hash(). #2870 (2.3.3)

Build/test system improvements and platform ports:

• CMake build system and scripts:
  • Instead of defaulting to looking for Python 2.7, the OIIO build now
    defaults to whatever Python is found (though a specific one can still
    be requested via the PYTHON_VERSION variable). #2705 (2.3.0.0/2.2.8)
    #2764 (2.3.0.1/2.2.8)
  • Make the OIIO CMake files work properly if OIIO is a subproject. Also
    various other CMake script refactoring. #2770 (2.3.1.1/2.2.9)
  • Extend checked_find_package with VERSION_MIN and VERSION_MAX #2773
    (2.3.1.1/2.2.9), DEFINITIONS and SETVARIABLES #3061 (2.3.7)
  • No longer directly link against python libraries when unnecessary.
    #2807 (2.2.11/2.3.3)
  • On Windows, fix some linkage problems by changing the pybind11
    bindings to make a CMake "shared" library rather than "module". Sounds
    wrong, but seems to work. We will reverse if this causes problems.
    #2830 (2.3.3/2.2.11)
  • Improvements to building or linking static libraries. #2854 (2.2.12/2.3.3)
  • Change default STOP_ON_WARNING to OFF for release branches (including
    this one) so that small change in compiler warnings after our release
    don't break anybody's builds. (Though we still stop on warnings for CI
    builds). #2861 (2.2.12/2.3.3)
  • The pkgconfig OpenImageIO.pc was specifying the include path
    incorrectly. #2869 (2.2.12/2.3.3)
  • Support for CMake 3.20. #2913 #2931 (2.3.4)
  • Be sure to have the namespace include the patch number for pre-release
    builds from the master branch. #2948 (2.3.4)
  • Use modern style cmake targets for PNG and ZLIB dependencies. #2957
    (2.3.4)
  • Propagate C++14 minimum requirement to downstream projects via our
    cmake config exports. #2965 (2.3.5)
  • Fix exported cmake config files, which could fail if Imath and OpenEXR
    weren't the at the same version number. #2975 (2.3.5/2.2.15)
  • Change default postfix for debug libraries to _d. (2.3.5)
  • Our CMake checked_find_package is now able to be requested to favor
    an exported config file, if present, on a package-by-package basis.
    #2984 (2.3.5/2.2.15)
  • If a package is requested to be disabled, skip its related tests rather
    than reporting them as broken. #2988 (2.3.5/2.2.16)
  • Better support for running testsuite when the build dir in odd places.
    #3065 #3067 (2.3.7.1)
  • To prevent accidental overwrite of sensitive areas (such as
    /usr/local), you now need to explicitly set CMAKE_INSTALL_PREFIX if you
    want the "install" to not be local to the build area. #3069 (2.3.7.1)
• Dependency version support:
  • C++20 is now supported. #2891 (2.3.3)
  • Fix deprecation warnings when building with very new PugiXML versions.
    #2733 (2.3.0.1/2.2.8)
  • Fixes to build against OpenColorIO 2.0. #2765 (2.3.0.1/2.2.8) #2817
    #2849 (2.3.3/2.2.11) #2911 (2.3.4)
  • Fix to accommodate upcoming OpenColorIO 2.1 deprecation of
    parseColorSpaceFromString. #2961 (2.3.5/2.2.15)
  • Work to ensure that OIIO will build correctly against Imath 3.0 and
    OpenEXR 3.0, as well as Imath 3.1 / OpenEXR 3.1. #2771 (2.3.1.1/2.2.9)
    #2876 #2678 #2883 #2894 (2.3.3/2.2.12) #2935 #2941 #2942 #2947 (2.3.4)
  • Better finding of OpenJpeg 2.4. #2829 (2.3.3/2.2.11)
  • On Mac, libheif 1.10 is very broken. Don't use that version. #2847
    (2.3.3/2.2.11)
  • Fix build break against changes coming in future libtiff, where it
    is changing from some libtiff-defined integer types to the equivalent
    stdint.h types. #2848 (2.3.3/2.2.11)
  • Fixes to support the libraw 202101 snapshot (their in-progress 0.21.0).
    #2850 (2.3.3/2.2.11)
  • More clear warnings about using OpenVDB 8+ when building for C++11,
    because OpenVDB 8 requires C++14 or higher. #2860 (2.2.12/2.3.3)
  • More gracefully handle building against a custom Imath/OpenEXR even
    when another exists in the system area. #2876 (2.2.12/2.3.3)
  • Minor fixes to build cleanly against the upcoming Imath 3.0. #2878
    (2.2.12/2.3.3)
  • Remove obsolete dependency on Boost random (now use std::random).
    #2896 (2.3.3)
  • libtiff 4.3 is supported. #2953 (2.3.4)
  • We now include a build_OpenJPEG.bash script that can conveniently
    build a missing OpenJPEG dependency. (2.3.5/2.2.16)
  • Changes to make it build against TBB 2021. #2985 (2.3.5/2.2.15)
  • Support for building OIIO with gcc 11. #2995 (2.2.16/2.3.6)
  • Fixes to accommodate Imath 3.1 upcoming changes. #2996 (2.2.16/2.3.6)
  • Finding FFMpeg now correctly detects the version. #2994 (2.2.16/2.3.6)
  • FFMpeg minimum version is now >= 3.0. #2999 (2.3.6)
  • Fixes for detecting and using Ptex, among other things got the version
    wrong. #3001 (2.2.16/2.3.6)
  • Fixes for building against fmt 8.0. #3007 (2.3.6)
  • The OpenCV minimum version is now >= 3.0. #3015 (2.3.6)
  • Fixes to build properly against the upcoming OpenColorIO 2.1. #3050
    (2.3.7)
  • Finding boost is more flexible when desiring static libraries. #3031
    (2.3.7/2.2.17)
  • FindTBB.cmake module updated to create proper targets. #3060 (2.3.7)
  • All the src/build_scripts/build_*.bash scripts now honor an env
    variable called DEP_DOWNLOAD_ONLY, which if set will only do the
    downloads but not the builds. #3058 #3072 (2.3.7)
  • Better finding of OpenCV on Windows. #3062 (2.3.7.1)
• Testing and Continuous integration (CI) systems:
  • Completely get rid of the old appveyor CI. #2782 (2.3.2.0)
  • Test against libtiff 4.2 in the "latest releases" test. #2792 (2.3.2.0)
  • Got Windows CI fully working, bit by bit. #2796 #2798 #2805 #2821 #2826
    #2834 #2835 #2836 #2838 #2839 #2840 (2.3.3)
  • Reduce redundant verbosty in CI output logs during dependency building.
  • Modify hash_test to verify correctness and stability of the hashes.
    #2853 (2.3.3)
  • When building custom OpenEXR with build_openexr.bash, don't have it
    build the examples. #2857 (2.2.12/2.3.3)
  • Speed up CI by using GitHub 'cache' actions + ccache. #2859 (2.2.12/2.3.3)
  • Separate stages (setup, deps, build, test) into separate GHA "steps"
    for better logging and understanding of the timing and performance.
    #2862 (2.2.12/2.3.3)
  • Now actively testing libheif in Linux CI. #2866 (2.2.12/2.3.3)
  • Remove the last vestiges of Travis-CI, which we no longer use. #2871
    (2.2.12/2.3.3)
  • For failed tests, add CMake cache and log part of the saved artifacts.
    (2.2.12/2.3.3)
  • CI now tests build in C++20 mode. #2891 (2.3.3)
  • Our CI clang-format test now uses LLVM/clang-format 11. #2966
  • Test the clang11 + C++17 combo. #3004 (2.3.6)
• Platform support:
  • Fixes for mingw. #2698 (2.3.0.0)
  • Windows fix: correct OIIO_API declaration on aligned_malloc,
    aligned_free of platform.h. #2701 (2.3.0.0)
  • Fix boost linkage problem on Windows. #2727 (2.3.0.1/2.2.8)
  • Fix warnings when compiling webpinput.cpp on 32 bit systems. #2783
    (2.3.2.0)
  • Fix problems with copysign sometimes defined as a preprocessor symbol
    on Windows. #2800 (2.3.2)
  • Fixes related to Mac M1-based systems: Fix crash in ustring internals
    #2990 (2.3.5/2.2.15.1)

Notable documentation changes:

• Make Readthedocs generate downloadable HTML as well as PDF. #2746
• Explain how to read image data into separate per-channel buffers. #2756
  (2.3.0.1/2.2.8)
• Improve use of breathe and formatting. #2762 (2.3.1.0)
• Remove documentation of deprecated ImageBufAlgo functions. #2762 (2.3.1.0)
• Document missing texture option fields. #2762 (2.3.1.0)
• Starting to use "sphinx_tabs" as a clear way to present how things should
  be done for different language bindings. #2768 (2.3.1.0)

Release 2.2.18 (1 Sep 2021) -- compared to 2.2.17

• Honor env variable CUE_THREADS (used by OpenCue) to set the default size
  of OIIO's thread pool. #3038
• Compatibility with OpenColorIO 2.1. #3050
• Dev: Extend Sysutil::getenv() to take a default if the environment variable
  is not found. #3047 #3048

At long last, I have branched for this year's big release, OpenImageIO 2.3. The new branch is "dev-2.3" (a change from the old convention of RB-x.y; I think it's clearer since many people didn't know what RB stood for), and the first beta is tagged "v2.3.7.0-beta". Why 2.3.7.0? Because we had 6 "developer preview" snapshots along the way.

I'm aiming for September 1, 2021 to tag a release. It might get delayed if serious problems are discovered. But considering that SPI and probably lots of other places use master branch or developer preview snapshots for real production, for anything other changes made in the last couple weeks, I'm pretty confident that everything is solid and has already been through the heat of battle in production. In general, the only difference between our release branches and master is that release branches make a strict promise about not breaking compatibility; there is little practical difference in stability or production worthiness.

We will strive to make no breaking changes to the API or ABI between now and the release. After final release, when the beta designation disappears, we will be 100% strict about this promise.

We will keep releasing 2.2.x patches as needed, but they will slow down and be restricted to only critical bug fixes, as we transition to 2.3 being the current supported release. Meanwhile, master -- which is now the staging area for what will become 2.4 next year -- is where big new, and possibly compatibility-breaking, changes will be made.

Release 2.3 (1 Sept 2021?) -- compared to 2.2

New minimum dependencies and compatibility changes:

• C++ standard: C++14 is now the minimum (gcc 6.1 - 11.2, clang 3.4 - 12,
  MSVS 2017 - 2019, icc 17+). The default C++ standard mode, if none is
  explicitly specified, is now C++14. #2918 (2.3.4) #2955 #2977 (2.3.5)
• FFMmpeg (optional) dependency minimum is now 3.0 (raised from 2.6). #2994
• OpenCV (optional) dependency minimum is now 3.0 (raised from 2.0). #3015

New major features and public API changes:

• Changes and clarifications to geterror() in all classes that have one:
  • geterror() now takes an optional clear parameter controls if the
    pending error is cleared as well as retrieved (default true).
  • All classes with geterror() are ensured to have has_error().
  • geterror() appends to the pending error message, if one is already
    present but not retrieved. Appending errors will be automatically
    separated by newlines, also any newline at the end of a geterror
    call will be stripped. #2740 (2.3.0.1)
• ImageInput error messages are now thread-specific. This prevents multiple
  threads simultaneously reading from the same ImageInput from getting the
  other thread's error messages. But it means you must always retrieve an
  error with getmessage() from the same thread that made the read call
  that generated the error. #2752 (2.3.0.1)
• ImageInput and ImageOutput have removed several fields (like the m_mutex)
  and instead now use a PIMPL idiom that hides internals better from the
  ABI. If you had a custom ImageInput/Output class that directly accessed
  the m_mutex, replace it instead with calls to ImageInput::lock() and
  unlock() (and ImageOutput). #2752 (2.3.1.0)
• Clarify that ImageBuf methods subimage(), nsubimages(), miplevel(),
  nmipevels(), and file_format_name() refer to the file that an ImageBuf
  was read from, and are thus only meaningful for ImageBuf's that directly
  read from files. "Computed" or copied ImageBuf's are solo images in memory
  that no longer correspond to a file even if the input to the computation
  was from a file. #2759 (2.3.0.1/2.2.9)
• New get_extension_map() returns a map that list all image file formats
  and their presumed file extensions. #2904 (2.3.4)
• Overhaul of how we handle "thumbnails" (small preview images embedded in
  the headers of some image file formats) #3021 (2.3.7):
  • Thumbnails no longer embedded in the ImageSpec as "thumbnail_image"
    metadata of a big character array blob.
  • ImageInput, ImageBuf, and ImageCache have varieties of get_thumbnail()
    method that returns the thumbnail as an ImageBuf. If this is not
    called, ideally the reader will not undergo the expense of reading and
    storing thumbnail data.
  • ImageOutput and ImageBuf have set_thumbnail() to communicate a
    thumbnail image as an ImageBuf.
  • The net effect is that reading files and retrieving ImageSpec's should
    be more efficient for apps that don't have a use for thumbnails.
• Significant improvements to UDIM handling in ImageCache/TextureSystem:
  • New UDIM texture name patterns recognized: %(UDIM)d is the Houdini
    convention, and _u##v## is for Animal Logic's internal renderer.
    #3006 (2.2.16/2.3.6)
  • It is now permissible for get_texture_info()/get_image_info() to
    retrieve metadata for UDIM patterns -- previously it failed, but now
    it will succeed in retrieving any metadata as long as it is present
    and has the same value in all the matching "UDIM tile" panels. #3049
    (2.3.7)
  • TextureSystem now exposes methods is_udim(), resolve_udim(),
    inventory_udim() that should be helpful for apps dealing with UDIM
    textures. See the documentation for details. #3054 (2.3.7)
  • Performance improvements when using UDIM textures. #3049 (2.3.7)
• ImageBuf, when "wrapping" an app-owned buffer, now allows explicit
  specification of stride lengths. This allows you to wrap a buffer that
  has internal padding or other non-contiguous spacing of pixels, scanlines,
  or image planes. #3022 (2.3.6)
• oiiotool new commands and options:
  • --pastemeta takes two images as arguments, and appends all the
    metadata (only) from the first image onto the second image's pixels
    and metadata, producing a combined image. #2708 (2.3.0.0)
  • --chappend and --siappend both allow an optional modifier :n=
    to specify the number of images from the stack to be combined
    (default n=2). #2709 (2.3.0.0)
  • --fit now takes an additional optional modifier: fillmode= with
    choices of letterbox (default), width, and height. #2784 (2.3.2.0)
  • --autocc now has an optional modifier :unpremult=1 that causes
    any color transformations related to autocc to unpremultiply by alpha
    before the transformation and then re-premultiply afterwards, exactly
    the same control that exists for individual --colorconvert but never
    existed for autocc. #2890 (2.3.3)
  • --list-formats lists all the image file formats that OIIO knows
    about, and for each, the associated file extensions. #2904 (2.3.4)
  • --skip-bad-frames causes an error (such as an input file not being
    found) to simply skip to the next frame in the frame range loop, rather
    that immediately exiting. #2905 (2.3.4)
  • When doing expression substitution, {getattribute(name)} will be
    replaced by the value that OIIO::getattribute(name, ...) would
    retrieve. #2932 (2.3.4)
  • --missingfile (which takes a subsequent argument of error, black,
    or checker) determines the behavior when an input image file is
    missing (the file does not exist at all, does not include cases where
    the file exists but there are read errors). The default value of
    error matches the old behavior: report an error and terminate all
    command processing. Other choices of black or checker will continue
    processing but substitute either a black or checkerboard image for the
    missing file. #2949 (2.3.4)
  • --printinfo prints verbose metadata info about the current top
    image on the stack. #3056 (2.3.7)
  • --printstats prints statistics about the current top image on the
    stack, and optionally can limit the statistics to a rectangular
    subregion of the image. #3056 (2.3.7)
  • Expanded expression now support the following new metadata notations:
    META, METABRIEF, and STATS. #3025 (2.3.6)
  • --mosaic now takes optional fit=WxH modifier that lets you set the
    "cell" size, with all constituent images resized as needed (similarly
    to if you used --fit on each individually). This is helpful for
    creating contact sheets without having to resize separately or to know
    the resolution of the original images. #3026 (2.3.6)
  • REMOVED --histogram which had been deprecated and undocumented since
    before OIIO 2.0. #3029 (2.3.6)
• Python bindings:
  • When transferring blocks of pixels (e.g., ImageInput.read_image()
    or ImageOutput.write_scanline()), "half" pixels ended up mangled
    into uint16, but now they use the correct numpy.float16 type. #2694
    (2.3.0.0)
  • The value passed to attribute(name, typedesc, value) can now be a
    tuple, list, numpy array, or scalar value. #2695 (2.3.0.0)
  • Added Python bindings for the TextureSystem. #2842
  • Add the previously (inadvertently) omitted enum value for
    ImageBufAlgo.MakeTxBumpWithSlopes. #2951 (2.3.4)
• Environment variables:
  • CUE_THREADS env variable, if set, is now honored to set the default
    size of the OIIO thread pool. This helps to make OIIO-based apps be
    good citizens when run as OpenCue jobs. #3038 (2.3.7)
• Library organization:
  • All the utility classes are now in libOpenImageIO_Util only and
    libOpenImageIO depends on and links to libOpenImageIO_Util, rather
    than the utility classes being defined separately in both libraries.
    #2906 (2.3.4)
  • Include file change: imagebuf.h no longer includes imagecache.h.
    It should never have needed to, since it didn't need any of its
    declarations. Other code that should always have included both headers,
    but inadvertently only included imagebuf.h, may now need to add an
    explicit #include <OpenImageIO/imagecache.h>. #3036 (2.3.7)
• Important change to Makefile wrapper: We have a 'Makefile' that just wraps
  cmake for ease of use. This has changed its practice of putting builds in
  build/ARCH to just build (and local installs from dist/ARCH to dist)
  to better match common cmake practice. So where your local builds end up may
  shift a bit. #3057 (2.3.7)
• Deprecations: A number of long-deprecated functions and methods have been
  given deprecation attributes that will cause warnings if you use them. They
  will eventually be removed entirely at the next "breaking" release (3.0,
  whenever that is). #3032 (2.3.7)

Performance improvements:

• Speed up BMP reading by eliminating wasteful allocation and copying
  done on each scanline read. #2934 (2.3.4)
• For IBA::to_OpenCV(), improve efficiency for certain image copies. #2944
  (2.3.4)
• Fix runaway parsing time for pathological XMP metadata. #2968 (2.3.5/2.2.15)
• TextureSystem reduction in overhead when using UDIM textures. #3049 (2.3.7)

Fixes and feature enhancements:

• Fix a situation where if an ImageBuf backed by an ImageCache reads an
  image, then the image changes on disk, then another ImageBuf or ImageCache
  tries to read it, it could end up with the old version. This involved some
  strategic cache invalidation when ImageBuf's write images to disk. #2696
  (2.3.0.0)
• Improve parsing of XMP records in metadata: more correct handling of
  lists/sequences, better inference of types that look like int or float
  (rather than forcing unknown fields into strings), fixed bugs in parsing
  rational values. #2865 (2.2.12/2.3.3)
• ImageBuf/ImageBufAlgo:
  • IBA::contrast_remap() fixes bug that could crash for very large
    images #2704 (2.3.0.0)
  • Fix stack overflow crash in IBA::colorconvert of unusually wide images.
    #2716 (2.3.0.1/2.2.8)
  • Fix IBA::make_texture() incorrectly setting tile sizes. #2737
    (2.3.0.1/2.2.8)
  • IBA::make_texture() now correctly propagates error messages, which
    can be retrieved via the global OIIO::geterror() call. #2747 (2.3.0.1)
  • IBA::fit() now takes a fillmode parameter that controls exactly
    how the resize will occur in cases where the aspect ratio of the new
    frame doesn't exactly match that of the source image. The old way (and
    the default value now) is "letterbox", but now also "width" and
    "height" modes are allowed. See fit() documentation for details.
    #2784 (2.3.2.0)
  • IBA::fit() fixes some slight ringing at image edges when using a
    combination of black wrap mode and a filter with negative lobes.
    #2787 (2.3.2.0)
  • Fix: ociolook() and ociofiletransform() internally reversed the
    order of their inverse and unpremult arguments, making it hard to
    select the inverse transformation. #2844 (2.3.3/2.2.11)
  • Fix crash for certain calls to ImageBuf::set_write_format() when
    writing files that support per-channel data types. #2885 (2.3.3)
  • Fix: IBA::fillholes_pushpull did not correctly understand which
    channel was alpha when generating subimages. #2939 (2.3.4)
  • Fix: IBA::render_text did not properly account for non-1 alpha of
    the text color. #2981 (2.3.5/2.2.15)
  • IBA::colorconvert, colormatrixtransform, ociolook, ociodisplay,
    and ociofiletransform have been fixed so that if input is more than
    4 channels, the additional channels will be copied unchanged, rather
    than inadvertently set to 0. #2987 (2.3.5/2.2.16)
  • Thread safety has been improved for ImageBuf::init_spec() and read()
    (though no problem was ever reported in the wild). #3018 (2.3.6)
  • render_text() now accepts text strings with embedded linefeeds and
    will turn them into multiple lines of rendered text. #3024 (2.3.6)
• ImageCache/TextureSystem/maketx:
  • Fix ImageCache bug: add_tile/get_tile not properly honoring when
    chend < chbegin it should get all channels. #2742 (2.3.0.1/2.2.8)
  • ImageBufAlgo::make_texture() (as well as maketx and oiiotool -otex)
    clamp half values to their maximum finite range to prevent very large
    float inputs turning into Inf when saved as half values. #2891 (2.3.3)
  • Fix crash when sampling non-zero-base channels. #2962 (2.3.5/2.2.15)
  • IBA::make_texture() rejects "deep" input images, since they cannot
    be made into textures. Doing this could previously crash. #2991
    (2.3.5/2.2.16)
  • maketx and oiiotool -otex have fixed double printing of error
    messages #2992 (2.3.5).
  • Making bumpslopes textures now allows scaled slopes UV normalization.
    This is exposed as new maketx argument --uvsopes_scale, or passing
    the attribute uvslopes_scale (int) to make_texture(). #3012 (2.3.6)
  • There is a new variety of ImageCache::invalidate() that takes an
    ImageHandle* argument instead of a filename. #3035 (2.3.7)
• oiiotool:
  • --resize of images with multi-subimages could crash. #2711 (2.3.0.0)
  • Improve oiiotool's guessing about the desired output format based on
    inputs (in the absence of -d to specify the format). #2717
    (2.3.0.1/2.2.8)
  • --text now accepts text strings with embedded linefeeds and
    will turn them into multiple lines of rendered text. #3024 (2.3.6)
• BMP
  • Fix reading BMP images with bottom-to-top row order. #2776
    (2.3.1.1/2.2.9)
  • Correctly read BMP images of the older V1 variety without crashing.
    #2899 (2.3.3/2.2.13)
  • Fix error in the rightmost pixel column when reading 4 bpp images with
    odd horizontal resolution. #2899 (2.3.3/2.2.13)
  • BMP reads now return metadata "bmp:bitsperpixel" (only when not 24
    or 32) to indicate non-whole-byte channels sizes in the file, and
    "bmp:version" to indicate the version of the BMP format. #2899
    (2.3.3/2.2.13)
  • Speed up BMP reading by eliminating wasteful allocation and copying
    done on each scanline read. #2934 (2.3.4)
  • For BMP files that are 8 bits per pixel and all palette entries have
    R == G == B values, read the file as an 8-bit grayscale image instead
    of needlessly promoting to a full RGB 3-channel image. #2943 (2.3.4)
  • Full support for reading RLE-compressed BMP images. #2976 (2.3.5/2.2.15)
  • Write single channel BMP as 8 bit palette images. #2976 (2.3.5/2.2.15)
• DPX
  • Output to DPX files now supports IOProxy. (Input already did.) #3013
    (2.2.17/2.3.6)
• FFMpeg/movies:
  • Avoid potential crash when a frame can't be read. #2693 (2.3.0.0)
  • Several varieties of encoded videos with 12 bits per channel and
    having alpha were incorrectly reported as 8 bits and without alpha.
    #2989 (2.3.5/2.2.16)
• GIF
  • Support UTF-8 filenames on Windows for GIF files. #2777 (2.3.2.0)
  • Fix error checking for non-existant GIF files. #2886 (2.3.3)
• HEIC
  • Fix error decoding Apple HEIF files. #2794/#2809 (2.3.2.0)
  • Better valid_file() check for HEIF. #2810 (2.3.2.0)
  • Enabled AVIF decoding of heic files (requires libheif >= 1.7 and for
    it to have been built with an AV1 encoder/decoder). #2811 #2812 #2814
    #2818 (2.3.3.0)
• IFF
  • Fix broken reads of 16 bit iff files. #2736 (2.3.0.1/2.2.8)
• JPEG
  • Allow reading of JPEG files with mildly corrupted headers. #2927 (2.3.4)
• OpenEXR:
  • Fix rare crash that was possible when multithreaded writing openexr
    files. #2781 (2.3.2.0)
  • Improved error reporting when errors are encountered writing OpenEXR
    files. #2783 (2.3.2.0)
  • Fix potential crash parsing OpenEXR header that contains Rational
    attributes with certain values. #2791 (2.2.10/2.3.2)
  • EXPERIMENTAL: When building against OpenEXR 3.1, the OIIO CMake option
    -DOIIO_USE_EXR_C_API=ON will use a new OpenEXR API that we think
    will allow higher performance texture reads in a multithreaded app.
    This has not yet been benchmarked or tested thoroughly. #3009 #3027
    (2.3.6)
  • The reader no longer renames file metadata "version" into the special
    "openexr:version" that would indicate the OpenEXR file format version
    of the file. That's never what it meant! It really was just arbitrary
    metadata. #3044 (2.3.7)
• PNG
  • Read Exif data from PNG files. #2767 (2.3.1.1/2.2.9)
• PSD
  • Add support for reading ISD profiles for PSD input. #2788 (2.3.2.0)
  • Fix loading PSB files with cinf tags. #2877 (2.2.12/2.3.2)
• RAW:
  • Additional input configuration hints to control options in the
    underlying LibRaw: "raw:user_flip" #2769 (2.3.1.0)
    "raw:balance_clamped", "raw:apply_scene_linear_scale",
    "raw:camera_to_scene_linear_scale" #3045 (2.3.7)
  • Correctly handle RAW files with Unicode filenames on Windows. #2888
    (2.3.3)
  • Fix garbled output when raw:Demosaic hint is "none". #3045 (2.3.7)
• Targa
  • Fix potential crash when reading files with no thumbnail. #2903
    (2.3.3/2.2.13)
  • Fix alpha handling for some files. #3019 (2.3.6)
• TIFF:
  • Fix broken reads of multi-subimage non-spectral files (such as
    photometric YCbCr mode). #2692 (2.3.0.0)
  • Fix spec() and spec_dimensions() for MIPmapped TIFF files, they did
    not recognize being asked to return the specs for invalid subimage
    indices. #2723 (2.3.0.1/2.2.7)
  • Add ability to output 1bpp TIFF files. #2722 (2.3.0.1/2.2.7)
  • Fix reading TIFF files with "separate" planarconfig and rowsperstrip
    more than 1. #2757 (2.3.0.1/2.2.9)
  • Fix incorrect reading of tiled TIFF files using certain rare TIFF
    features that also have a vertical resolution that is not a whole
    multiple of the tile size. #2895 (2.3.3/2.2.13)
  • Support IOProxy for reading TIFF files. #2921 (2.3.4)
  • TIFF plugin now properly honors caller request for single-threaded
    operation. #3016 (2.3.6)
• WebP:
  • Add support for requesting compression "lossless". #2726 (2.3.0.1/2.2.8)
  • Input improvements including: RGB images are kept as RGB instead of
    always adding alpha; more efficient by skipping alpha premultiplication
    when it's unnecessary; now can read animated WebP images as
    multi-subimage files. #2730 (2.3.0.1/2.2.8)
• Fix memory leak during decoding of some invalid Exif blocks. #2824
• Fix possible divide-by-zero error in read_image/read_scanlines for invalid
  image sizes (usually from corrupt files). #2983 (2.3.5/2.2.16)

Developer goodies / internals:

• Internals now use C++11 final keywords wherever applicable. #2734
  (2.3.0.1)
• More internals conversion of old Strutil::sprintf to Strutil::fmt::format
  and related changes. #2889 (2.3.3)
• Redundant format conversion code removed from imageio.cpp #2907 (2.3.4)
• Eliminate direct output to std::cerr and std::cout by library calls.
  #2923 (2.3.4)
• argparse.h:
  • ArgParse::abort() lets the response to a command line argument signal
    that no further arguments should be parsed. #2820 (2.3.3/2.2.11)
• color.h:
  • New ColorConfig::OpenColorIO_version_hex() returns the hex code for
    the version of OCIO we are using (0 for no OCIO support). #2849 (2.3.3)
• farmhash.h:
  • Clean up all non-namespaced preprocessor symbols that are set
    by this header and may pollute the caller's symbols. #3002 (2.2.16/2.3.6)
• filesystem.h:
  • New Filesystem::generic_filepath() returnss a filepath in generic
    format (not OS specific). #2819 (2.3.3/2.2.11)
  • Improve exception safety in Filesystem directory iteration. #2998
    (2.2.16/2.3.6)
  • New filename_to_regex() makes a filename "safe" to use as a
    regex pattern (for example, properly backslashing any . characters).
    #3046 (2.3.7)
  • Some methods that used to take a const std::string& now take a
    string_view. #3047 (2.3.7)
• fmath.h:
  • Use CPU intrinsics to speed up swap_ending (by 8-15x when swapping
    bytes of large arrays). #2763 (2.3.1.0)
• hash.h:
  • farmhash::inlined::Hash now is constexpr and works for Cuda.
    #2843 (2.3.3) #2914 #2930 (2.3.4)
• oiioversion.h:
  • New macros OIIO_VERSION_GREATER_EQUAL and OIIO_VERSION_LESS.
    #2831 (2.3.3/2.2.11)
• platform.h:
  • New macro OIIO_INLINE_CONSTEXPR, equivalent to inline constexpr for
    C++17, but just constexpr for C++ <= 14. #2832 (2.3.3/2.2.11)
• simd.h:
  • Fix incorrect ARM NEON code in simd.h. #2739 (2.3.0.1/2.2.8)
• span.h:
  • std::size() and std::ssize() should work with OIIO::span now.
    #2827 (2.3.3/2.2.11)
• string_view.h:
  • std::size() and std::ssize() should work with OIIO::string_view
    now. #2827 (2.3.3/2.2.11)
  • More thorough constexr of string_view methods. #2841 (2.3.3)
• strongparam.h:
  • New StrongParam helper for disambiguating parameters. #2735 (2.3.2)
• strutil.h:
  • Strutil splits() and splitsv() should return no pieces when passed
    an empty string. (It previously erroneously returned one piece
    consisting of an empty string.) #2712 (2.3.0.0)
  • Fix build break when strutil.h was included in Cuda 10.1 code. #2743
    (2.3.0.1/2.2.8)
  • strhash() is now constexpr for C++14 and higher. #2843 (2.3.3)
  • New Strutil functions: find, rfind, ifind, irfind #2960 (2.3.4/2.2.14)
    contains_any_char #3034 (2.3.7) rcontains, ircontains #3053 (2.3.7)
• sysutil.h:
  • Extend getenv() to take a default if the environment variable is not
    found. #3037 #3040 (2.3.7)
• typedesc.h:
  • TypeDesc::basetype_merge(a,b) returns a BASETYPE having the
    precision and range to hold the basetypes of either a or b.
    #2715 (2.3.0.0)
  • TypeDesc can now describe 2D and 3D bounding boxes, as arrays of 2
    VEC2 aggregates (for 2D) or VEC3 aggregates (for 3D) with "BOX"
    semantic. The shorthand for these are TypeBox2, TypeBox3 (for
    float), and TypeBox2i and TypeBox3i for integer or pixel coordinte
    boxes. #3008 (2.2.17/2.3.6)
• unordered_map_concurrent.h:
  • New methods find_or_insert, nobin_mask(). #2867 (2.2.12/2.3.3)
• ustring.h:
  • ustring internals now have a way to ask for the list of ustrings whose
    hashses collided. #2786 (2.2.11/2.3.2.0)
  • ustring now guarantees that no two ustrings will return the exact same
    value for hash(). #2870 (2.3.3)

Build/test system improvements and platform ports:

• CMake build system and scripts:
  • Instead of defaulting to looking for Python 2.7, the OIIO build now
    defaults to whatever Python is found (though a specific one can still
    be requested via the PYTHON_VERSION variable). #2705 (2.3.0.0/2.2.8)
    #2764 (2.3.0.1/2.2.8)
  • Make the OIIO CMake files work properly if OIIO is a subproject. Also
    various other CMake script refactoring. #2770 (2.3.1.1/2.2.9)
  • Extend checked_find_package with VERSION_MIN and VERSION_MAX #2773
    (2.3.1.1/2.2.9), DEFINITIONS and SETVARIABLES #3061 (2.3.7)
  • No longer directly link against python libraries when unnecessary.
    #2807 (2.2.11/2.3.3)
  • On Windows, fix some linkage problems by changing the pybind11
    bindings to make a CMake "shared" library rather than "module". Sounds
    wrong, but seems to work. We will reverse if this causes problems.
    #2830 (2.3.3/2.2.11)
  • Improvements to building or linking static libraries. #2854 (2.2.12/2.3.3)
  • Change default STOP_ON_WARNING to OFF for release branches (including
    this one) so that small change in compiler warnings after our release
    don't break anybody's builds. (Though we still stop on warnings for CI
    builds). #2861 (2.2.12/2.3.3)
  • The pkgconfig OpenImageIO.pc was specifying the include path
    incorrectly. #2869 (2.2.12/2.3.3)
  • Support for CMake 3.20. #2913 #2931 (2.3.4)
  • Be sure to have the namespace include the patch number for pre-release
    builds from the master branch. #2948 (2.3.4)
  • Use modern style cmake targets for PNG and ZLIB dependencies. #2957
    (2.3.4)
  • Propagate C++14 minimum requirement to downstream projects via our
    cmake config exports. #2965 (2.3.5)
  • Fix exported cmake config files, which could fail if Imath and OpenEXR
    weren't the at the same version number. #2975 (2.3.5/2.2.15)
  • Change default postfix for debug libraries to _d. (2.3.5)
  • Our CMake checked_find_package is now able to be requested to favor
    an exported config file, if present, on a package-by-package basis.
    #2984 (2.3.5/2.2.15)
  • If a package is requested to be disabled, skip its related tests rather
    than reporting them as broken. #2988 (2.3.5/2.2.16)
  • Finding boost is more flexible when desiring static libraries. #3031
    (2.3.7/2.2.17)
  • FindTBB.cmake module updated to create proper targets. #3060 (2.3.7)
• Dependency version support:
  • C++20 is now supported. #2891 (2.3.3)
  • Fix deprecation warnings when building with very new PugiXML versions.
    #2733 (2.3.0.1/2.2.8)
  • Fixes to build against OpenColorIO 2.0. #2765 (2.3.0.1/2.2.8) #2817
    #2849 (2.3.3/2.2.11) #2911 (2.3.4)
  • Fix to accommodate upcoming OpenColorIO 2.1 deprecation of
    parseColorSpaceFromString. #2961 (2.3.5/2.2.15)
  • Work to ensure that OIIO will build correctly against Imath 3.0 and
    OpenEXR 3.0, as well as Imath 3.1 / OpenEXR 3.1. #2771 (2.3.1.1/2.2.9)
    #2876 #2678 #2883 #2894 (2.3.3/2.2.12) #2935 #2941 #2942 #2947 (2.3.4)
  • Better finding of OpenJpeg 2.4. #2829 (2.3.3/2.2.11)
  • On Mac, libheif 1.10 is very broken. Don't use that version. #2847
    (2.3.3/2.2.11)
  • Fix build break against changes coming in future libtiff, where it
    is changing from some libtiff-defined integer types to the equivalent
    stdint.h types. #2848 (2.3.3/2.2.11)
  • Fixes to support the libraw 202101 snapshot (their in-progress 0.21.0).
    #2850 (2.3.3/2.2.11)
  • More clear warnings about using OpenVDB 8+ when building for C++11,
    because OpenVDB 8 requires C++14 or higher. #2860 (2.2.12/2.3.3)
  • More gracefully handle building against a custom Imath/OpenEXR even
    when another exists in the system area. #2876 (2.2.12/2.3.3)
  • Minor fixes to build cleanly against the upcoming Imath 3.0. #2878
    (2.2.12/2.3.3)
  • Remove obsolete dependency on Boost random (now use std::random).
    #2896 (2.3.3)
  • libtiff 4.3 is supported. #2953 (2.3.4)
  • We now include a build_OpenJPEG.bash script that can conveniently
    build a missing OpenJPEG dependency. (2.3.5/2.2.16)
  • Changes to make it build against TBB 2021. #2985 (2.3.5/2.2.15)
  • Support for building OIIO with gcc 11. #2995 (2.2.16/2.3.6)
  • Fixes to accommodate Imath 3.1 upcoming changes. #2996 (2.2.16/2.3.6)
  • Finding FFMpeg now correctly detects the version. #2994 (2.2.16/2.3.6)
  • FFMpeg minimum version is now >= 3.0. #2999 (2.3.6)
  • Fixes for detecting and using Ptex, among other things got the version
    wrong. #3001 (2.2.16/2.3.6)
  • Fixes for building against fmt 8.0. #3007 (2.3.6)
  • The OpenCV minimum version is now >= 3.0. #3015 (2.3.6)
  • Fixes to build properly against the upcoming OpenColorIO 2.1. #3050
    (2.3.7)
  • All the src/build_scripts/build_*.bash scripts now honor an env
    variable called DOWNLOAD_ONLY, which if set will only do the downloads
    but not the builds. #3058 (2.3.7)
• Testing and Continuous integration (CI) systems:
  • Completely get rid of the old appveyor CI. #2782 (2.3.2.0)
  • Test against libtiff 4.2 in the "latest releases" test. #2792 (2.3.2.0)
  • Got Windows CI fully working, bit by bit. #2796 #2798 #2805 #2821 #2826
    #2834 #2835 #2836 #2838 #2839 #2840 (2.3.3)
  • Reduce redundant verbosty in CI output logs during dependency building.
  • Modify hash_test to verify correctness and stability of the hashes.
    #2853 (2.3.3)
  • When building custom OpenEXR with build_openexr.bash, don't have it
    build the examples. #2857 (2.2.12/2.3.3)
  • Speed up CI by using GitHub 'cache' actions + ccache. #2859 (2.2.12/2.3.3)
  • Separate stages (setup, deps, build, test) into separate GHA "steps"
    for better logging and understanding of the timing and performance.
    #2862 (2.2.12/2.3.3)
  • Now actively testing libheif in Linux CI. #2866 (2.2.12/2.3.3)
  • Remove the last vestiges of Travis-CI, which we no longer use. #2871
    (2.2.12/2.3.3)
  • For failed tests, add CMake cache and log part of the saved artifacts.
    (2.2.12/2.3.3)
  • CI now tests build in C++20 mode. #2891 (2.3.3)
  • Our CI clang-format test now uses LLVM/clang-format 11. #2966
  • Test the clang11 + C++17 combo. #3004 (2.3.6)
• Platform support:
  • Fixes for mingw. #2698 (2.3.0.0)
  • Windows fix: correct OIIO_API declaration on aligned_malloc,
    aligned_free of platform.h. #2701 (2.3.0.0)
  • Fix boost linkage problem on Windows. #2727 (2.3.0.1/2.2.8)
  • Fix warnings when compiling webpinput.cpp on 32 bit systems. #2783
    (2.3.2.0)
  • Fix problems with copysign sometimes defined as a preprocessor symbol
    on Windows. #2800 (2.3.2)
  • Fixes related to Mac M1-based systems: Fix crash in ustring internals
    #2990 (2.3.5/2.2.15.1)

Notable documentation changes:

• Make Readthedocs generate downloadable HTML as well as PDF. #2746
• Explain how to read image data into separate per-channel buffers. #2756
  (2.3.0.1/2.2.8)
• Improve use of breathe and formatting. #2762 (2.3.1.0)
• Remove documentation of deprecated ImageBufAlgo functions. #2762 (2.3.1.0)
• Document missing texture option fields. #2762 (2.3.1.0)
• Starting to use "sphinx_tabs" as a clear way to present how things should
  be done for different language bindings. #2768 (2.3.1.0)

Release 2.2.17 (1 Aug 2021) -- compared to 2.2.16

• Output to DPX files now supports IOProxy. (Input already did.) #3013
• typedesc.h: TypeDesc can now describe 2D and 3D bounding boxes, as arrays
  of 2 VEC2 aggregates (for 2D) or VEC3 aggregates (for 3D) with "BOX"
  semantic. The shorthand for these are TypeBox2, TypeBox3 (for float),
  and TypeBox2i and TypeBox3i for integer or pixel coordinte
  boxes. #3008
• Build: Fixes for building against fmt 8.0.0. #3007
• Build: Finding boost is more flexible when desiring static libraries. #3031

Release 2.2.16 (1 Jul 2021) -- compared to 2.2.15

• New UDIM texture name patterns recognized: %(UDIM)d is the Houdini
  convention, and _u##v## is for Animal Logic's internal renderer. #3006
  (2.2.16)
• When doing color space transforms on images with > 4 channels -- the
  additional channels are now copied unaltered, rather than leaving them
  black. #2987 (2.2.16)
• FFMpeg: fix some encodings that didn't correctly recognize that they were
  more than 8 bits, or had alpha. #2989 (2.2.16)
• farmhash.h: Clean up all non-namespaced preprocessor symbols that are set
  by this header and may pollute the caller's symbols. #3002 (2.2.16)
• Fix crashes on M1 (ARM) based Mac. #2990 (2.2.16)
• Bug fix: avoid divide-by-0 error computing chunk size for invalid image
  sizes. #2983 (2.2.16)
• make_texture (and maketx and oiiotool -otex) no longer crash if you
  try to make a texture out of a "deep" image; instead it will return an
  error message. #2991 (2.2.16)
• filesystem.h: Improve exception safety in Filesystem directory iteration.
  #2998 (2.2.16)
• Build: Improve finding of OpenJPEG. #2979 (2.2.16)
• Build: Support for building OIIO with gcc 11. #2995 (2.2.16)
• Build: Fixes to accommodate Imath 3.1 upcoming changes. #2996 (2.2.16)
• Build: Finding FFMpeg now correctly detects the version. #2994 (2.2.16)
• Build: clang + C++17 + LibRaw < 0.20 are mutually incompatible. Detect
  this combination and warn / disable libraw under those conditions. #3003
  (2.2.16)
• Build: Fix CMake behavior for REQUIRED_DEPS due to a typo. #3011 (2.2.16)
• Build: Fixes for detecting and using Ptex, among other things got the
  version wrong. #3001 (2.2.16)
• Testing: If a feature is disabled, skip its tests rather than reporting
  them as broken. #2988 (2.2.16)
• CI: Test the combination of clang and C++17. #3003 (2.2.16)

Release 2.2.15.1 (3 Jun 2021) -- compared to 2.2.15.0

• Fix crash / misbehavior in ustring internals on Apple M1 ARM. #2990