Research On Video Algorithm Based On Intel Xeon Phi Many-core Architecture

Frame rate up conversion(Frame Rate Up-Conversion,FRUC)technology came into being through the middle frame insertion in the original two video frames in a way that will improve the realization of video from a low frame rate to a higher frame rate.With the emergence of ultra high definition 4K video,the number of image pixels increase rapidly,resulting in that traditional serial algorithm is difficult to meet the requirements of fast processing or real-time processing.In recent years,heterogeneous many-core devices have become a important equipment in the field of high performance computing.In 2012,Intel released Intel Xeon Phi coprocessor which supports X86 instruction architecture.The device has more than 57 cores and computing architecture based on the X86.Each core has a 512 bit vector processing unit(Vector Processor Unit,VPU),and can handle 16 single precision floating-point arithmetic.At the same time,the coprocessor supports a variety of parallel programming languages,such as Open MP,Pthread,Open CL,etc.So this thesis proposes a parallel frame rate up conversion algorithm based on Intel Xeon Phi,including motion estimation algorithm and motion compensation algorithm.The main work in this thesis has the following two aspects:(1)This thesis proposes a video frame rate up conversion algorithm based on Open CL.Firstly,we analyze the motion estimation module,and we propose a parallel method based on the macroblock level for the 3DRS algorithm which has the characteristics of mutual dependence of the space candidate motion vectors.First,the image is divided into 240×135 blocks,we use the advantages of MIC multithreading,the parallel operation of the 240 blocks per row,that is,each core calculate a block of the candidate motion vector.Then,we carried out parallel analysis on motion compensation.Each pixel is independent in motion compensation module.The experimental results show that the Open CL based parallel scheme obtains good speedup.(2)This thesis proposes an algorithm of video frame rate conversion based on dual MIC card.The traditional frame rate conversion algorithm divides the video into a plurality of frames,each frame is processed by a serial motion estimation module and a motion compensation module,and all the algorithms are executed in sequence.We designed an asynchronous processing scheme,the motion estimation and motion compensation module for processing at the same time,because the nodes of experimental environment with two MIC cards,so we will put the motion estimation module in a MIC card,motion compensation module in a card on.The first MIC card calculates the motion vector data to be transmitted back to the host side,and then transfers the motion vector data to the second MIC cards from the host side,where in the data copy of the MIC card and the host side is controlled by the semaphore.