[x265] [ANN] x265 0.4.1

Steve Borho steve at borho.org
Wed Sep 25 07:27:59 CEST 2013


Hello,

I've made a tag on the stable branch for 0.4.1.  This email describes the
state of the encoder at that tag.

x265 can be compiled and run on Linux, Mac OS X, and Windows.  Its cmake
based build system supports MSVC 9-11, Xcode, gcc+gmake, MinGW/MSYS, and
Intel C++ (both icl and icpc)

x265 currently generates HM11 compliant bitstreams.

= New features since 0.3 =

1. Frame parallelism

  * The GOP level parallelism and mini-gop cadence of the 0.3 release
    were removed and replaced with fine-grained frame level parallelism.
  * The memory requirements were drastically reduced, x265 no longer
    caches subpel planes per reference frame
  * Latency is drastically reduced, WPP + 5 frame threads is enough to
    saturate a dual socket 8 core Xeon.

2. New simple average bitrate (ABR) rate control

  * converges on target bitrate via frame level QP adjustments
  * algorithm was adapted from x264 for HEVC
  * safe for use in combination with frame parallelsim (new since 0.4)

3. New Lookahead

  * three complexity levels of slice decision with scene cut and flash
detection
  * algorithm was adapted from x264 for HEVC
  * still missing bidir cost estimates, multiple refs, weightp, and MBtree


= Disabled Features since 0.3 =

  * weightp and weighted bidir prediction are both disabled; they were
    broken when we stopped pre-generating reference (subpel) pixels
  * GOP parallelism is dead, hopefully to never return to the core
    library.  It could be added above x265 if necessary.
  * many CLI options were added, removed, or renamed since 0.3.  Consult
    the online help for the current set of options.


= Known Bugs =

  * all intra encodes (--keyint 1) require --b-adapt 0 to avoid
    a lookahead bug
  * --no-sao-lcu-bounds cannot be used in conjunction with frame
    parallelism.  The encoder will use incomplete reference pixels
    with predictable bad effects.
  * we have a report that our DPB signalling is potentially incorrect
    for decoders which respect max DPB size. This is under investigation.

ABR and lookahead are very new and we will not be surprised to find bugs in
those features.  Please report any bugs you find to this mailing list.


= Performance Characteristics =

See https://bitbucket.org/multicoreware/x265/wiki/Performance


= Upcoming features =

We are trying to make the encoder deterministic with -Fn for all values of
n greater than 1.  In other words, there will always be a slight benefit to
having only one frame thread (unlimited motion search to reference frames)
but -F2 and -F10 should output the exact same bitstream with CQP rate
control.  This is considered a debugging feature to validate our frame
parallelism.

We are also planning to swap out the HM's bidir search with logic adapted
from x264 (for performance reasons) and to add estimation logic for this
new bidir search to the lookahead (to improve lookahead accuracy).

We plan to repair the weighted prediction features and re-enable its
command line options.

Lastly, the lookahead will use wave-front scheduling of the lowres CUs
using the existing thread pool.  This should lesson the bottlenecking
effects of --b-adapt 2.

-- 
Steve Borho
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20130925/32d2fdd1/attachment.html>


More information about the x265-devel mailing list