[FFmpeg-devel] GSoC 2021

[Guo, Yejun]

> -----Original Message-----
> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of
> Michael Niedermayer
> Sent: 2021年2月2日 17:48
> To: FFmpeg development discussions and patches <ffmpeg-devel at ffmpeg.org>
> Subject: [FFmpeg-devel] GSoC 2021
> 
> Hi all
> 
> Most people probably already know but just to be sure everyone knows GSoC
> 2021 is 175h not 350h
> https://groups.google.com/g/google-summer-of-code-discuss/c/GgvbLrFBcUQ
> ?pli=1
> 
> Some project ideas may need to be adjusted accordingly

hi, for your information.

just like last year, to get more chances for GSoC project, I proposed my idea at https://01.org/linuxmedia/news/gsoc-2021-ideas, and also copy here for your convenience.

Async support for TensorFlow backend in FFmpeg
Description: FFmpeg DNN (deep neural network) module is to enable deep learning based FFmpeg filters with OpenVINO, TensorFlow and Native as its backends. The OpenVINO backend has supported async+batch execution for model inference with commit e67b5d and 64ea15 etc, it shows good performance gain. This project focuses on the async support for TensorFlow backend which invokes TensorFlow C library for model loading and inference. The main difference between the two backends is that OpenVINO API has async execution function while TensorFlow C API does not expose async function, so we need to support async execution by multiple threads within TensorFlow backend with sync function TF_SessionRun. (it is a plus if batch mode is also supported.)
Difficulty: Medium
Skill Required: C, multiple threads, DNN knowledge, git