Recherche avancée

Sur d’autres sites (8266)

• ffmpeg as subprocess takes too long

  12 décembre 2019, par almosthero

  One part of my program requires FFMPEG to transcode files using GPU.

  When I execute the FFMPEG command alone in terminal, without Python, it takes aproximetly 0.7 second. Running 50 simultaneous commands doesn’t affect the results. The time is always around 0.7 sec (thanks to a powerful GPU that can handle that amount of requests).

  Putting this command to subprocess increases the time to 1.7 second and raises up to 20 seconds when I run multiple instance even though this command should work asynchronously.

  def run_ffmpeg(filename, gop, representation, output_filename=None):
  
      process_result = subprocess.run(
  
          args=[
  
              'sudo', FFMPEG_PATH,
  
              '-vsync', '0',
  
              '-hwaccel', 'cuvid', # transport decoded file through GPU without touching CPU
  
              '-hwaccel_device', '0',
  
              '-c:v', 'h264_cuvid', # Decoding using GPU
  
              '-i', filename,
  
              # ...
  
              '-c:v', 'hevc_nvenc', # Encoding using GPU
  
              # ...more filters...
  
              output_filename
  
          ]
  
      )
  
      log_process_result(process_result)
  
      return output_filename

  The sequence is way more complex :

  def transcode_file(media_id, target_repr, target_nr):
  
      # ... variables definitions
  
      run_ffmpeg(target_segment_path, gop='120', representation=target_repr, output_filename=result)
  
  
  
  def prepare_file(media_id, representation_id, file_name):
  
      # ... variables definitions
  
          path = transcode_file(media_id, target_repr, target_nr)
  
  
  
  async def download_media_handler(request):
  
      # ... variables definitions
  
      file_info = await loop.run_in_executor(
  
          executor, prepare_file, media_dir_name, representation_id, file_name)

  Neither GPU or CPU is fully used so this is not the case.

  Do you know why ? Does subprocess really takes so much time to run ?

  How can I rewrite this to minimize the time ? Are there any quicker alternatives ?

• Force x264 to only use long term reference frames

  12 décembre 2019, par cloudraven

  I want to generate a h.264 stream in which P-frames are using references exclusively using long-term references identifiers rather than defaulting to a short-term reference given by the previous frame.

  At this point this is an academic experiment to understand the encoding behavior of long-term references and the consequences of forcing frames to refer to a specific frame rather than several.

  It would be great if I could do this from the command line, but if not possible I am open to modifying x264 for this purpose.

  Could anyone point me to where in the source code it is determined :

  • the reference frame / macroblock / partition for the current macroblock
  • changes made to the short term, long term reference frame list
  • whether to use a short term reference frame or a long term reference frame.

  Using another encoder like nvenc is also acceptable.

• Force x264 to only use long term reference frames

  12 décembre 2019, par cloudraven

  I want to generate a h.264 stream in which P-frames are using references exclusively using long-term references identifiers rather than defaulting to a short-term reference given by the previous frame.

  At this point this is an academic experiment to understand the encoding behavior of long-term references and the consequences of forcing frames to refer to a specific frame rather than several.

  It would be great if I could do this from the command line, but if not possible I am open to modifying x264 for this purpose.

  Could anyone point me to where in the source code it is determined :

  • the reference frame / macroblock / partition for the current macroblock
  • changes made to the short term, long term reference frame list
  • whether to use a short term reference frame or a long term reference frame.

  Using another encoder like nvenc is also acceptable.