Recherche avancée

Sur d’autres sites (7100)

• avfilter/vf_removegrain : replace qsort with AV_QSORT

  25 octobre 2015, par Ganesh Ajjanagadde

  avfilter/vf_removegrain : replace qsort with AV_QSORT
  filter_slice calls qsort, so qsort is in a performance critical
  
  position. AV_QSORT is substantially faster due to the inlining of the
  
  comparison callback. Thus, the increase in performance is worth the
  
  increase in binary size.
  Sample benchmark (x86-64, Haswell, GNU/Linux),
  
  filter-removegrain-mode-02 (from FATE)
  
  new :
  
    24060 decicycles in qsort,       1 runs,      0 skips
  
    15690 decicycles in qsort,       2 runs,      0 skips
  
     9307 decicycles in qsort,       4 runs,      0 skips
  
     5572 decicycles in qsort,       8 runs,      0 skips
  
     3485 decicycles in qsort,      16 runs,      0 skips
  
     2517 decicycles in qsort,      32 runs,      0 skips
  
     1979 decicycles in qsort,      64 runs,      0 skips
  
     1911 decicycles in qsort,     128 runs,      0 skips
  
     1568 decicycles in qsort,     256 runs,      0 skips
  
     1596 decicycles in qsort,     512 runs,      0 skips
  
     1614 decicycles in qsort,    1024 runs,      0 skips
  
     1874 decicycles in qsort,    2046 runs,      2 skips
  
     2186 decicycles in qsort,    4094 runs,      2 skips
  old :
  
   246960 decicycles in qsort,       1 runs,      0 skips
  
   135765 decicycles in qsort,       2 runs,      0 skips
  
    70920 decicycles in qsort,       4 runs,      0 skips
  
    37710 decicycles in qsort,       8 runs,      0 skips
  
    20831 decicycles in qsort,      16 runs,      0 skips
  
    12225 decicycles in qsort,      32 runs,      0 skips
  
     8083 decicycles in qsort,      64 runs,      0 skips
  
     6270 decicycles in qsort,     128 runs,      0 skips
  
     5321 decicycles in qsort,     256 runs,      0 skips
  
     4860 decicycles in qsort,     512 runs,      0 skips
  
     4424 decicycles in qsort,    1024 runs,      0 skips
  
     4191 decicycles in qsort,    2046 runs,      2 skips
  
     4934 decicycles in qsort,    4094 runs,      2 skips
  Reviewed-by : Michael Niedermayer <michael@niedermayer.cc>
  
  Signed-off-by : Ganesh Ajjanagadde <gajjanagadde@gmail.com>
  

  • [DH] libavfilter/vf_removegrain.c

• avcodec/huffman : replace qsort with AV_QSORT

  23 octobre 2015, par Ganesh Ajjanagadde

  avcodec/huffman : replace qsort with AV_QSORT
  ff_huff_build_tree uses qsort underneath. AV_QSORT is substantially
  
  faster due to the inlining of the comparison callback. Furthermore, this
  
  code is reasonably performance critical, since in e.g the fraps codec,
  
  ff_huff_build_tree is called on every frame. This routine is also called
  
  in vp6 on every frame in some circumstances.
  Sample benchmark (x86-64, Haswell, GNU/Linux), vp6 from FATE :
  
  vp6 (old) :
  
    78930 decicycles in qsort,       1 runs,      0 skips
  
    45330 decicycles in qsort,       2 runs,      0 skips
  
    27825 decicycles in qsort,       4 runs,      0 skips
  
    17471 decicycles in qsort,       8 runs,      0 skips
  
    12296 decicycles in qsort,      16 runs,      0 skips
  
     9554 decicycles in qsort,      32 runs,      0 skips
  
     8404 decicycles in qsort,      64 runs,      0 skips
  
     7405 decicycles in qsort,     128 runs,      0 skips
  
     6740 decicycles in qsort,     256 runs,      0 skips
  
     7540 decicycles in qsort,     512 runs,      0 skips
  
     9498 decicycles in qsort,    1024 runs,      0 skips
  
     9938 decicycles in qsort,    2048 runs,      0 skips
  
     8043 decicycles in qsort,    4095 runs,      1 skips
  vp6 (new) :
  
    15880 decicycles in qsort,       1 runs,      0 skips
  
    10730 decicycles in qsort,       2 runs,      0 skips
  
    10155 decicycles in qsort,       4 runs,      0 skips
  
     7805 decicycles in qsort,       8 runs,      0 skips
  
     6883 decicycles in qsort,      16 runs,      0 skips
  
     6305 decicycles in qsort,      32 runs,      0 skips
  
     5854 decicycles in qsort,      64 runs,      0 skips
  
     5152 decicycles in qsort,     128 runs,      0 skips
  
     4452 decicycles in qsort,     256 runs,      0 skips
  
     4161 decicycles in qsort,     511 runs,      1 skips
  
     4081 decicycles in qsort,    1023 runs,      1 skips
  
     4072 decicycles in qsort,    2047 runs,      1 skips
  
     4004 decicycles in qsort,    4095 runs,      1 skips
  Reviewed-by : Timothy Gu <timothygu99@gmail.com>
  
  Reviewed-by : Michael Niedermayer <michael@niedermayer.cc>
  
  Signed-off-by : Ganesh Ajjanagadde <gajjanagadde@gmail.com>
  

  • [DH] libavcodec/huffman.c

• avutil/mathematics : speed up av_gcd by using Stein’s binary GCD algorithm

  11 octobre 2015, par Ganesh Ajjanagadde

  avutil/mathematics : speed up av_gcd by using Stein’s binary GCD algorithm
  This uses Stein’s binary GCD algorithm :
  
  https://en.wikipedia.org/wiki/Binary_GCD_algorithm
  
  to get a roughly 4x speedup over Euclidean GCD on standard architectures
  
  with a compiler intrinsic for ctzll, and a roughly 2x speedup otherwise.
  
  At the moment, the compiler intrinsic is used on GCC and Clang due to
  
  its easy availability.
  Quick note regarding overflow : yes, subtractions on int64_t can, but the
  
  llabs takes care of that. The llabs is also guaranteed to be safe, with
  
  no annoying INT64_MIN business since INT64_MIN being a power of 2, is
  
  shifted down before being sent to llabs.
  The binary GCD needs ff_ctzll, an extension of ff_ctz for long long (int64_t). On
  
  GCC, this is provided by a built-in. On Microsoft, there is a
  
  BitScanForward64 analog of BitScanForward that should work ; but I can’t confirm.
  
  Apparently it is not available on 32 bit builds ; so this may or may not
  
  work correctly. On Intel, per the documentation there is only an
  
  intrinsic for _bit_scan_forward and people have posted on forums
  
  regarding _bit_scan_forward64, but often their documentation is
  
  woeful. Again, I don’t have it, so I can’t test.
  As such, to be safe, for now only the GCC/Clang intrinsic is added, the rest
  
  use a compiled version based on the De-Bruijn method of Leiserson et al :
  
  http://supertech.csail.mit.edu/papers/debruijn.pdf.
  Tested with FATE, sample benchmark (x86-64, GCC 5.2.0, Haswell)
  
  with a START_TIMER and STOP_TIMER in libavutil/rationsl.c, followed by a
  
  make fate.
  aac-am00_88.err :
  
  builtin :
  
  714 decicycles in av_gcd,    4095 runs,      1 skips
  de-bruijn :
  
  1440 decicycles in av_gcd,    4096 runs,      0 skips
  previous :
  
  2889 decicycles in av_gcd,    4096 runs,      0 skips
  Signed-off-by : Ganesh Ajjanagadde <gajjanagadde@gmail.com>
  
  Signed-off-by : Michael Niedermayer <michael@niedermayer.cc>
  

  • [DH] libavutil/intmath.h
  • [DH] libavutil/mathematics.c