HysiaDecode

The sub-project of Hysia-T2O. Please go to the Hysia-T2O to see the overview of the project.

This repository aims to provide a highly efficient and user-friendly video preprocessing library.

The whole pipeline of the HysiaDecode can be illustrated in the following diagram. The Decode module firstly detect the GPU exist. If GPU exists, the module will select the GPU decoding function priorly. Otherwise, the module will select the CPU decoding function. It depends on user‘s hardware configuration.So it support various types of users. After decoding, we save the data into a queue, and the applications can fetch the frames from the queue. To facilitate Python users, we warp the C++ function to Python using Pybind11. You can follow the example you provided in test directory. Enjoy your experience and don't forget to report the issue you encountered!

pATH	DESCRIPTION
include	header files(include the dependencies)
lib	required libraries(E.g., OpenCV, FFmpeg, Cuvid)
src	implementation of decoding function
python	warp the C++ function to Python
test	test files
build	save the generated files(.o and .so) after make
utils	some pre-implemented functions in NvDecode

Required Packages and Platform

Requirement	Minimal Version
Platform	Ubutun 16/14.04, Manjaro 17.10 etc.
Nvidia GPU	Titan X, M6000 etc.
Driver	nvidia-396 or later
CUDA	CUDA 8.0 or later

Installation

Before using our repository, you should install Nvidia-driver first. For ubuntu user, you can install it under this instruction belowing:

sudo apt-get purge nvidia*
sudo add-apt-repository ppa:graphics-drivers 
sudo apt-get update
sudo apt-get install nvidia-396 
lsmod | grep nvidia 

Pybind11 is also necessary, because we use Pybind wrap the C++ function to Python

pip install pybind11

Then:

git clone https://github.com/iversonicter/HysiaDecode.git

cd HysiaDecode

If the required hardware or driver is not available:

make CPU_ONLY=TRUE

Otherwise:

make

After compliation and link, this reporistory will generate a shared file like this 'PyDecoder.cpython-3xm-x86_64-linux-gnu.so' in build directory. You can import it in Python like this, don't forget add it into the system path unless you will meet this error " no moudle named PyDecoder ":

import PyDecoder

Prefermance

we test the decoding speed on our server(Titan X, Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz) using several video clips from The Big Bang Series. We also compare the decoding speed with released algorithms in OpenCV. We achieved the considerable performance. Need to notice, our library provide Python interface which can be easily used for users who don't familiar with C/C++. This is only a alpha version of decoding. In the next step, we will continue optimizing the library to achieve a large throughput. You are also welcomed to join our team to contribute your code. We also found that GPU initialization takes up much time. If you want to decode a short video clip, we recommend using CPU. Otherwise, GPU is preferred.

Hardware	Speed
GPU decoding(Titan X)	~800 frames/s
CPU (Intel E5-2630)	~200 frames/s
Python cv2.VideoCapture	375 frames /s
OpenCV VideoCaptur*	399 frames /s
OpenCV cv::cudacodec::VideoReader*	865 frames/s

Issues

Some issues we met in develop and test. If you meet the same error, you can refer the solution we provide here.

• issue 1

GPU in use: Quadro M6000
Traceback (most recent call last):
  File "test_gpu_decoder.py", line 19, in <module>
    dec.ingestVideo(video_path)
RuntimeError: NvDecoder : cuvidCtxLockCreate(&m_ctxLock, cuContext) returned error -1282253296 at src/NvDecoder/NvDecoder.cpp:524

Solution : This error comes from cuvid link conflicts. You should link

• Issue 2

the decoding processing is blocked and does not start the decoded process

Solution: When it happens, maybe your GPU does support NVDecode. You should try to consult the Nvidia documents to check whether you graphical card are suitable. In this repository, we detached the decoded function from Video_Codec_SDK_8.2.16.

Reference

We referred the released repositories listed below:

Nvidia video_codec_SDK_8.2.16

scanner video decoding

Nvidia NVVL

FFmpeg official tutorials