NEWS

Movit 1.6.2, March 18th, 2018

  - Various bugfixes.


Movit 1.6.1, January 31st, 2018

  - Loosen up some restrictions on strong one-to-one-effects. In particular,
    this fixes a bug with MixEffect and compute shaders that hit some
    Nageru setups.


Movit 1.6.0, January 24th, 2018

  - Support for effects that work as compute shaders. Compute shaders are
    generally slower than fragment shaders for the same algorithm,
    but allow some forms of communication between shader invocations
    and have more flexible output, which can enable more efficient algorithms. 
    See effect.h for more details. Note that the fastest rendering API on
    EffectChain is now to a texture if possible, not to an FBO. This will
    only matter if the last effect is a compute shader.
    
  - Movit now includes a compute shader implementation of DeinterlaceEffect,
    which is automatically used instead of the fragment shader implementation
    if your GPU and OpenGL driver supports it (in practice, this means on
    all platforms except on macOS). The compute shader version is typically
    20–80% faster than the fragment shader version, depending on your GPU
    and other factors.

    A compute shader implementation of ResampleEffect was written but
    ultimately failed to be faster, and so is not included.

  - Support for microbenchmarks of effects through the Google microbenchmarking
    framework (optional). Currently, DeinterlaceEffect and ResampleEffect has
    benchmarks; enable them by running the unit test with --benchmark (also try
    --benchmark --help).

  - Effects can now explicitly request _not_ to have mipmaps, which means they
    can do so without needing to request bounce and fiddling with the sampler
    state. Note that this is an API change for effects.

  - Movit now requires C++11, both to build and to #include the header files.
    Support for SDL1 has been dropped; unit tests and the demo program now need
    SDL2.

  - Various smaller bugfixes and optimizations.
    

Movit 1.5.3, August 10th, 2017

  - Various bugfixes.


Movit 1.5.2, July 5th, 2017

  - The texture generation for ResampleEffect (when the size is changed)
    now uses less CPU.

  - Work around an NVIDIA driver threading bug by aggressively caching VAOs.


Movit 1.5.1, May 29th, 2017

  - Allow YCbCrInput to change input format after finalize.

  - Some minor YCbCrInput bugfixes.


Movit 1.5.0, March 21st, 2017

  - Support interleaved Y'CbCr input (4:4:4 in a single texture).

  - Support 10-bit and 12-bit Y'CbCr, both for input and output. For planar,
    these are supported packed in 16-bit ints; for interleaved, 10:10:10:2 is
    supported. (Efficient conversion to and from v210, ie. 10-bit 4:2:2,
    is possible using compute shaders, but Movit does not include support
    for them at the current point.) Note that this now means the num_levels
    flag in YCbCrFormat actually matters, although 0 will be interpreted
    as 256 (8-bit) for the benefit of older applications.

  - Limited support for having multiple Y'CbCr outputs from a chain.

  - Allow changing the Y'CbCr output coefficients runtime, ie., after finalize.

  - Fix an issue where the last pass would have been rendered with the sRGB
    flag set, which confused Qt applications running in certain NVIDIA
    configurations.


Movit 1.4.0, November 5th, 2016

  - Allow setting the intermediate format for chains, instead of hardcoding
    it at 16-bit RGBA; advanced users can use this to e.g. ask for 8-bit
    sRGB intermediates, reducing the amount of memory bandwidth needed
    at the cost of reduced precision. Whether this is a good tradeoff or not
    depends on the exact chain and your requirements.

  - Fix an issue where a (cached) shader program could be used from multiple
    threads at a time, causing the uniforms to contain unpredictable values.

  - Make the error printed on check_error() slightly friendlier: Include the
    enum if possible, and print it to stderr instead of stdout.


Movit 1.3.2, February 23rd, 2016

  - Fix an issue with initialization in certain locales. Patch from
    Jean-Baptiste Mardelle.


Movit 1.3.1, February 15th, 2016

  - Fix an issue where certain effect chains (particularly involving
    out-of-tree effects that return only a constant color) could cause
    texture coordinates not to be set properly. (The new code probably
    also has slightly lower OpenGL driver overhead.) Reported by
    Christophe Thommeret.


Movit 1.3.0, January 31st, 2016

  - Movit now requires GLSL 1.30 (so a driver from 2008 or later);
    before, it claimed to support 1.10, but actually used 1.30 features.
    Note that some OpenGL drivers, in particular on OS X, only supports
    GLSL 1.30 (actually, 1.50) if you have an OpenGL core context.

  - Add a deinterlacer, based on YADIF.

  - Allow parallel output to RGBA and Y'CbCr textures at the same time.

  - Make FlatInput and YCbCrInput support taking in external OpenGL
    textures. Also allow them to change width/height after instantiation.

  - Various compatibility and performance fixes. In particular,
    fp16 conversion on older (non-Haswell) CPUs is much faster, due to
    new conversion code by Fabian Giesen.


Movit 1.2.0, September 24th, 2015

  - Movit now ships a version.h with a #define MOVIT_VERSION that
    increases on every API change, including in git. (The number
    itself carries no semantic meaning beyond this.) Movit 1.2.0
    is not API or ABI compatible with 1.1.x, so there has been a
    soname bump to libmovit.so.3.

  - More flexible Y'CbCr input; in particular, a special input
    for 4:2:2 interleaved data (UYVY), and support for semi-planar
    input (Cb and Cr in the same texture, like in NV12). Note that
    you should now also set the new “num_levels” member of
    YCbYCbCrFormat to 256 (signifying 8-bit input), although
    it is not used yet; this is a stepping stone towards supporting
    10- and 12-bit Y'CbCr.

  - Basic support for Y'CbCr _output_. Currently only 8-bit,
    and only 4:4:4 (ie., you'll need to subsample yourself
    afterwards). It is possible to split the output into multiple
    textures, though, if you want a luma/chroma split (NV12-like)
    or full planar.

  - Support top-left origin for output. This is only really useful
    if you are rendering directly into some memory area with top-left
    origin; most users will get the expected behavior by using
    bottom-left as before.

  - Rework uniform handling for less OpenGL overhead. Note that
    this means your effects now need to register uniforms in the
    C++ code instead of declaring them in the .frag file.

  - Make the PaddingEffect border subpixel-aware, and also support
    an arbitrary (subpixel) border offset. This means that you can
    compose PaddingEffect with an integral left/top offset
    (use the new IntegralPaddingEffect for potentially more speed)
    and ResampleEffect with a subpixel left/top offset (and zoom to
    compensate) to get Lanczos interpolation for the pixels and a
    simple bilinear interpolation for the border itself.

  - Fix a bug that could cause very bad filter weights in
    ResampleEffect.

  - Various performance improvements, in particular with regards to
    CPU usage in ResampleEffect.


Movit 1.1.3, March 29th, 2015

  - Fix accuracy issues in ResampleEffect, particularly when
    zooming. Note that this might cost a few percent performance.
    Reported by Christophe Thommeret.

  - Make the number of BlurEffect taps configurable.
    Based on patch by Christophe Thommeret.

  - Multiple other bugfixes, some courtesy Dan Dennedy and
    Christophe Thommeret. In particular, thread-safety fixes
    related to locales and chain finalization.


Movit 1.1.2, August 12th, 2014

  - Performance bugfix: Fix texture freelist behavior so that it's
    LRU instead of the exact opposite. Patch by Christophe Thommeret.

  - Performance bugfix: Correct the number of blur taps read in the
    blur filter (it was reading about twice as many as it should).
    Also found by Christophe Thommeret.


Movit 1.1.1, April 12th, 2014

 - Fix an issue that could cause assertion failure in ResourcePool
   when FBOs are reused, especially with NVIDIA's drivers.


Movit 1.1, April 10th, 2014

 - NOTE: Movit now uses libepoxy rather than GLEW. This is needed because
   of core context and GLES3 support (see below). Movit 1.1 should be
   generally API-compatible with 1.0.3 (with the exception of the next
   point), but is not ABI-compatible, so there has been a soname bump
   to libmovit.so.2.

 - NOTE: The rules for using an EffectChain or ResourcePool in multiple
   OpenGL contexts or threads have changed somewhat; see the comments in
   resource_pool.h.

 - Movit will now work in OpenGL core contexts, and on GLES 3.0 or newer.
   (GLES2 devices are not supported.) This also holds for all unit tests,
   provided that you use SDL2 and not SDL1 (because SDL2 is needed to set
   up such contexts). Note that the included demo application still requires
   a classic OpenGL context.

 - ResampleEffect can now do sub-pixel translate and/or zoom.

 - LumaMixEffect now has an “invert” flag, as a convenience to e.g. change
   a left-to-right wipe into a right-to-left one.

 - Significant reduction in driver overhead, especially on NVIDIA drivers.

 - Various smaller bugfixes and performance improvements.


Movit 1.0.3, March 16th, 2014

 - Yet more build system tweaks mainly related to distribution packaging.


Movit 1.0.2, March 16th, 2014

 - Make a few tweaks to “make install”, to make distributions' lives easier.


Movit 1.0.1, March 16th, 2014

 - Fix so that shared libraries are built.


Movit 1.0, March 16th, 2014

 - Initial release.