Recherche avancée

Sur d’autres sites (8190)

• FFMPEG and AWS : What's the most efficient way to handle this ?

  28 mai 2022, par Red Vic

  I'm new to AWS and I originally built the FFmpeg functions on my Node.JS API. But I realized this is the wrong way to do it in a real-world app, and that you need to use separate Lambda functions in AWS that handle the video editing separately from the main server.

  


  I'm mainly a front-end developer but I'm open to learning new things.

  


  I basically have the following process in my app :

  


  

  • User uploads video.

  


  • I need to take that video and add a watermark to it.

  


  • I then need a copy of the watermarked video in a smaller resolution.

  


  • I then need a 6 seconds GIF of the smaller resolution video.

  


  • Finally, I need to upload the 3 edited files (2 .mp4's and 1 .gif) to S3, and remove the original, non-watermarked video.

  


  


  Here are my questions to be clear :

  


  

  • Should I upload the original file to S3 or to the server ? And why ?

  


  • Is the process above doable in a single Lambda function ? Or do I need more Lambda functions ?

  


  • How would you handle this problem, personally ?

  


  


  I originally built it by chaining one function to the next with promises, but AWS seems like a different world of doing things and the way I originally built it would not work.

  


  Thanks a lot.

  


  Update
Here are some tests I did with a couple videos :

  


  





  


  


  


  


  


  


  


  


  



  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  


  	Test 1	Test 2	Test 3	Test 4	Test 5
  Original video resolution	1080p	1080p	1080p	1080p	480p
  Original video duration	23 minutes	15 minutes	11 minutes	3.5 minutes	5 minutes
  Step 1 duration (Watermarking original video)	30 minutes	18 minutes	14 minutes	4 minutes	2 minutes
  Step 2 duration (Watermarking lower resolution)	5 minutes	3 minutes	3 minutes	1 minute	skip (already low res)
  Step 3 duration (6 seconds GIF creation)	negligible (15 seconds)	negligible (10 seconds)	negligible (7 seconds)	negligible	negligible
  Total	35 minutes	21 minutes	17 minutes	5 minutes	2 minutes

  


• lavfi/nlmeans : make compute_safe_ssd_integral_image_c faster

  6 mai 2018, par Clément Bœsch

  lavfi/nlmeans : make compute_safe_ssd_integral_image_c faster
  before :  ssd_integral_image_c : 49204.6
  
  after :   ssd_integral_image_c : 44272.8
  Unrolling by 4 made the biggest difference on odroid-c2 (aarch64) ;
  
  unrolling by 2 or 8 both raised 46k cycles vs 44k for 4.
  Additionally, this is a much better reference when writing SIMD (SIMD
  
  vectorization will just target 16 instead of 4).
  

  • [DH] libavfilter/vf_nlmeans.c

• hardware conversion of image pixel format in ffmpeg ?

  18 janvier 2024, par dongrixinyu

  I am trying to decode an online rtmp video stream into RGB format frames, and then encoding RGB frames into an online stream.

  


  Task

  


  Here is what I do now :

  


  


  decoding a video stream to get images(RGB) ---> ai model process ---> encoding frames(RGB) to form a video stream in H264

  


  


  My scheme

  


  All my code in written in C with FFmpeg dependencies. The detailed steps are :

  


  


  rtmp/rtsp video stream ---> AVPacket ---(nvidia cuda)---> AVFrame(nv12 pix fmt) ---> AVFrame(RGB pix fmt) ---> AI process.

  


  


  


  AVFrame(RGB pix fmt) ---> AVFrame(nv12 pix fmt) ---(nvidia cuda)---> AVPacket ---> rtmp/rtsp video stream

  


  


  Now, the decoding and encoding part are run on NVIDIA GPU, which is quite fast.

  


  But the conversion of pixel format between AV_PIX_FMT_NV12 and AV_PIX_FMT_RGB is run on CPU, which is astonishingly CPU-consuming cause the size of video frame is 2k.

  


  My question

  


  So, is there any off-the-shelf method to fulfill the conversion of image pixel format on GPU (especially via cuda) directly ?