ffmpeg-hardware-decode

Use ffmpeg and NVCodec to hardware decode video or video stream. in DNN inference system

Decoder

 Video.mp4   ->   cuvid(GPU Decoder)   ->   GPU Memory  ->   Normialize Cuda Kernel   ->   
 float*  Channel  based   ->   DNN Inference

Input Format: Supported File/USB Camera/RTSP/RTMP/CustomStream etc.

Normialize Kernel: Subtract mean and divide std, convert YUV to BGR or RGB

Encoder

 Camera/Other Source   ->   nvenc(GPU Encoder)   ->   Custom Memory Stream/File

Output Format: Only supported mp4

Environment

• ffmpeg 4.2

• nasm 2.14.02

• nv-codec-headers

• opencv 4.2.0 [optional]

• nvcodec-VideoCodecSDK 10.0.26

• cuda 10.2

• x264-snapshot-20190704-2245-stable

• GPU 2080Ti [optional]

Or download all lean code ffmpeg.hw.lean.code.tar.gz

Startup

• note: /datav/newbb/lean is root directory
  • ​ /datav/newbb/lean/build lean code directory
  • ​ /datav/newbb/lean/lean lean build result binary directory

Change the directory in *_build file, to your path

> mkdir build && mkdir lean
> cd build
> tar -zxvf ../ffmpeg.hw.lean.code.tar.gz
> mv Video_Codec_SDK_10.0.26 ../lean/

• Change /datav/newbb/lean to your path, after to run this code

> bash nasm_build
> bash nvcodec_build
> bash x264_build
> bash ffmpeg_build
> bash opencv_build

Run examples

> make run -j8