Recherche avancée

Sur d’autres sites (8915)

• lavc/cbrt_tablegen : speed up tablegen

  11 janvier 2016, par Ganesh Ajjanagadde

  lavc/cbrt_tablegen : speed up tablegen
  This exploits an approach based on the sieve of Eratosthenes, a popular
  
  method for generating prime numbers.
  Tables are identical to previous ones.
  Tested with FATE with/without —enable-hardcoded-tables.
  Sample benchmark (Haswell, GNU/Linux+gcc) :
  
  prev :
  
  7860100 decicycles in cbrt_tableinit,       1 runs,      0 skips
  
  7777490 decicycles in cbrt_tableinit,       2 runs,      0 skips
  
  [...]
  
  7582339 decicycles in cbrt_tableinit,     256 runs,      0 skips
  
  7563556 decicycles in cbrt_tableinit,     512 runs,      0 skips
  new :
  
  2099480 decicycles in cbrt_tableinit,       1 runs,      0 skips
  
  2044470 decicycles in cbrt_tableinit,       2 runs,      0 skips
  
  [...]
  
  1796544 decicycles in cbrt_tableinit,     256 runs,      0 skips
  
  1791631 decicycles in cbrt_tableinit,     512 runs,      0 skips
  Both small and large run count given as this is called once so small run
  
  count may give a better picture, small numbers are fairly consistent,
  
  and there is a consistent downward trend from small to large runs,
  
  at which point it stabilizes to a new value.
  Reviewed-by : Michael Niedermayer <michael@niedermayer.cc>
  
  Signed-off-by : Ganesh Ajjanagadde <gajjanagadde@gmail.com>
  

  • [DH] libavcodec/cbrt_tablegen.h

• Further SMC Encoding Work

  25 août 2011, par Multimedia Mike — General

  Sometimes, when I don’t feel like doing anything else, I look at that Apple SMC video encoder again.

  8-bit Encoding
  When I last worked on the encoder, I couldn’t get the 8-color mode working correctly, even though the similar 2- and 4-color modes were working fine. I chalked the problem up to the extreme weirdness in the packing method unique to the 8-color mode. Remarkably, I had that logic correct the first time around. The real problem turned out to be with the 8-color cache and it was due to the vagaries of 64-bit math in C. Bit shifting an unsigned 8-bit quantity implicitly results in a signed 32-bit quantity, or so I discovered.

  Anyway, the 8-color encoding works correctly, thus shaving a few more bytes off the encoding size.

  Encoding Scheme Oddities
  The next step is to encode runs of data. This is where I noticed some algorithmic oddities in the scheme that I never really noticed before. There are 1-, 2-, 4-, 8-, and 16-color modes. Each mode allows encoding from 1-256 blocks of that same encoding. For example, the byte sequence :

    0x62 0x45
  

  Specifies that the next 3 4×4 blocks are encoded with single-color mode (of byte 0×62, high nibble is encoding mode and low nibble is count-1 blocks) and the palette color to be used is 0×45. Further, opcode 0×70 is the same except the following byte allows for specifying more than 16 (i.e., up to 256) blocks shall be encoded in the same matter. In light of this repeat functionality being built into the rendering opcodes, I’m puzzled by the existence of the repeat block opcodes. There are opcodes to repeat the prior block up to 256 times, and there are opcodes to repeat the prior pair of blocks up to 256 times.

  So my quandary is : What would the repeat opcodes be used for ? I hacked the FFmpeg / Libav SMC decoder to output a histogram of which opcodes are used. The repeat pair opcodes are never seen. However, the single-repeat opcodes are used a few times.

  Puzzle Solved ?
  I’m glad I wrote this post. Just as I was about to hit “Publish”, I think I figured it out. I haven’t mentioned the skip opcodes yet– there are opcodes that specify that 1-256 4×4 blocks are unchanged from the previous frame. Conceivably, a block could be unchanged from the previous frame and then repeated 1-256 times from there.

  That’s something I hadn’t thought of up to this point for my proposed algorithm and will require a little more work.

  Further reading

  • My first entry regarding my toy SMC encoder

• Merge commit ’d371c3c2e2830d9783465ecfe1ab7d93351083b7’

  17 mai 2013, par Michael Niedermayer

  Merge commit ’d371c3c2e2830d9783465ecfe1ab7d93351083b7’
  * commit ’d371c3c2e2830d9783465ecfe1ab7d93351083b7’ :
  
    vf_frei0r : make config_props work properly when called multiple times.
  
    vf_gradfun : make config_props work properly when called multiple times.
  
    vf_lut : make config_props work properly when called multiple times.
  Conflicts :
  
  	libavfilter/vf_lut.c
  Merged-by : Michael Niedermayer <michaelni@gmx.at>
  

  • [DH] libavfilter/vf_frei0r.c
  • [DH] libavfilter/vf_gradfun.c
  • [DH] libavfilter/vf_lut.c