Recherche avancée

Sur d’autres sites (5976)

• Android ImageFormat.YUV_420_888 in FFmpeg

  5 janvier 2020, par Yondonator

  Im using ffmpeg on Android. I have an image with the format ImageFormat.YUV_420_888and want to put the image into an AVFrame. The Problem is that I dont know what format ImageFormat.YUV_420_888 is in ffmpeg. I tried with AV_PIX_FMT_YUV420P but the result looks like this :
  Wrong result

  It shoud look so :
  Right result

  Here is how I copy the data :

  Image.Plane[] planes = image.getPlanes();
  
  for(int i = 0; i < planes.length; i++)
  
  {
  
      BytePointer plane = new BytePointer(planes[i].getBuffer());
  
      tmpAVFrame.data(i, plane);
  
  }

  Edit :
  With Pixel format AV_PIX_FMT_YUV440P only the colors are wrong :
  Wrong Colors

• Converting MSB padding to LSB padding

  23 novembre 2020, par Oier Lauzirika Zarrabeitia

  I am using FFmpeg and Vulkan for a video related project. The problem is that FFmpeg uses MSB padding for the 9, 10 and 12bit formats, whilst Vulkan accepts LSB padded pixel formats.

  


  I am using the following code to convert from one another, but the performance is terrible (1 second per frame), which cannot be accepted for video playback.

  


  if(paddingBits > 0) {
    //As padding bits are set to 0, using CPU words for the shifting should not affect the result
    assert(dstBuffers[plane].size() % sizeof(size_t) == 0);
    for(size_t i = 0; i < dstBuffers[plane].size(); i += sizeof(size_t)) {
        reinterpret_cast(dstBuffers[plane][i]) <<= paddingBits;
    }
}


  


  Note that this code will be executed 1-4 times per frame (once per plane) and dstBuffers[plane].size() will be in the order of a couple of MegaBytes. dstBuffers[plane]'s data has an alignment greater than size_t's, so I am not performing unaligned read/writes. Using a smaller type such as uint16_t makes it perform worse.

  


  My question is if there is any standard function (or inside the FFmpeg library) which implements the same behavior in a fancy way so that the performance is not a concern. If not, is there a more efficient way to implement it ?

  


  Edit :dstBuffers[plane] stores std::byte-s and although is not of type std::vector, (it is a custom class) it is contiguous.

  


• lavfi/hqdn3d : add slice thread optimization

  2 octobre 2019, par Jun Zhao

  lavfi/hqdn3d : add slice thread optimization
  Enabled one thread per plane, used the test command for 1080P video
  
  (YUV420P format) as follow :
  ffmpeg -i 1080p.mp4 -an -vf hqdn3d -f null /dev/nul
  This optimization improved the performance about 30% in 1080P YUV420P
  
  case (from 110fps to 143fps), also pass the framemd5 check and FATE.
  Reviewed-by : Paul B Mahol <onemda@gmail.com>
  
  Reviewed-by : Moritz Barsnick <barsnick@gmx.net>
  
  Signed-off-by : Jun Zhao <barryjzhao@tencent.com>
  

  • [DH] libavfilter/vf_hqdn3d.c
  • [DH] libavfilter/vf_hqdn3d.h