2011-03-07 v0.9.6 "Bali"
  Our second named release, focused on a faster, higher quality, encoder.

  - Upgrading:
    This release is backwards compatible with Aylesbury (v0.9.5). Users
    of older releases should refer to the Upgrading notes in this
    document for that release.

  - Enhancements:
      vpxenc --psnr shows a summary when encode completes
      --tune=ssim option to enable activity masking
      improved postproc visualizations for development
      updated support for Apple iOS to SDK 4.2
      query decoder to determine which reference frames were updated
      implemented error tracking in the decoder
      fix pipe support on windows

  - Speed:
      Primary focus was on good quality mode, speed 0. Average improvement
      on x86 about 40%, up to 100% on user-generated content at that speed.
      Best quality mode speed improved 35%, and realtime speed 10-20%. This
      release also saw significant improvement in realtime encoding speed
      on ARM platforms.

        Improved encoder threading
        Dont pick encoder filter level when loopfilter is disabled.
        Avoid double copying of key frames into alt and golden buffer
        FDCT optimizations.
        x86 sse2 temporal filter
        SSSE3 version of fast quantizer
        vp8_rd_pick_best_mbsegmentation code restructure
        Adjusted breakout RD for SPLITMV
        Changed segmentation check order
        Improved rd_pick_intra4x4block
        Adds armv6 optimized variance calculation
        ARMv6 optimized sad16x16
        ARMv6 optimized half pixel variance calculations
        Full search SAD function optimization in SSE4.1
        Improve MV prediction accuracy to achieve performance gain
        Improve MV prediction in vp8_pick_inter_mode() for speed>3

  - Quality:
      Best quality mode improved PSNR 6.3%, and SSIM 6.1%. This release
      also includes support for "activity masking," which greatly improves
      SSIM at the expense of PSNR. For now, this feature is available with
      the --tune=ssim option. Further experimentation in this area
      is ongoing. This release also introduces a new rate control mode
      called "CQ," which changes the allocation of bits within a clip to
      the sections where they will have the most visual impact.

        Tuning for the more exact quantizer.
        Relax rate control for last few frames
        CQ Mode
        Limit key frame quantizer for forced key frames.
        KF/GF Pulsing
        Add simple version of activity masking.
        make rdmult adaptive for intra in quantizer RDO
        cap the best quantizer for 2nd order DC
        change the threshold of DC check for encode breakout

  - Bug Fixes:
      Fix crash on Sparc Solaris.
      Fix counter of fixed keyframe distance
      ARNR filter pointer update bug fix
      Fixed use of motion percentage in KF/GF group calc
      Changed condition for using RD in Intra Mode
      Fix encoder real-time only configuration.
      Fix ARM encoder crash with multiple token partitions
      Fixed bug first cluster timecode of webm file is wrong.
      Fixed various encoder bugs with odd-sized images
      vp8e_get_preview fixed when spatial resampling enabled
      quantizer: fix assertion in fast quantizer path
      Allocate source buffers to be multiples of 16
      Fix for manual Golden frame frequency
      Fix drastic undershoot in long form content