基于OpenCL机器视觉算法GPU实现
|
摘要
针对不断增长的对机器视觉算法处理效率和实时性的要求,研究基于异构编程框架OpenCL对机器视觉算法在通用计算机图形处理单元(GPU)上的并行处理和加速方法,提出结合存储分配、指令流优化、数据重用等方法的并行优化策略。在Sobel边缘检测、Canny边缘检测、Harris角点检测、高斯图像金字塔4个不同并行度视觉算法上进行验证,验证结果表明,在不考虑数据传输的情况下,对比CPU串行实现取得了平均6.16的加速比,对比OpenCV的GPU库(即CUDA实现)取得了1.12-5.47的加速比,验证了所提优化策略的有效性。
Aiming at the increasing demands for the processing efficiency and real-time performance of the machine vision algorithm,the parallel processing and accelerating method of machine vision algorithm based on heterogeneous programming framework open computing language(OpenCL)on computer graphics processing unit(GPU)was studied,and a parallel optimization strategy based on storage allocation,instruction flow optimization and data reuse was proposed.It was validated by four different degrees of parallelism vision algorithms including Sobel edge detection,Canny edge detection,Harris corner detection and Gaussian image pyramid.The results show that,without considering the data transmission,the average 6.16 acceleration ratio is obtained by comparing the central processing unit(CPU)serial implementation,and acceleration ratio of the 1.12-5.47 is obtained by comparing GPU library of OpenCV(i.e.the CUDA implementation),which verifies the effectiveness of the proposed optimization strategy.
Aiming at the increasing demands for the processing efficiency and real-time performance of the machine vision algorithm,the parallel processing and accelerating method of machine vision algorithm based on heterogeneous programming framework open computing language(OpenCL)on computer graphics processing unit(GPU)was studied,and a parallel optimization strategy based on storage allocation,instruction flow optimization and data reuse was proposed.It was validated by four different degrees of parallelism vision algorithms including Sobel edge detection,Canny edge detection,Harris corner detection and Gaussian image pyramid.The results show that,without considering the data transmission,the average 6.16 acceleration ratio is obtained by comparing the central processing unit(CPU)serial implementation,and acceleration ratio of the 1.12-5.47 is obtained by comparing GPU library of OpenCV(i.e.the CUDA implementation),which verifies the effectiveness of the proposed optimization strategy.
引文
[1]Sparsh M,Jeffrey SV.A survey of CPU-GPU heterogeneous computing techniques[J].ACM Computing Surveys,2015:47(4):69.
[2]Jain A,Namdev A,Chawla M.Parallel edge detection by SOBEL algorithm using CUDA C[C]//IEEE Students’Conference on Electrical,Electronics and Computer Science.Bhopal:IEEE,2016:1-6.
[3]Huang Y,Bai Y,Li R,et al.Research of Canny edge detection algorithm on embedded CPU and GPU heterogeneous systems[C]//International Conference on Natural Computation,Fuzzy Systems and Knowledge Discovery.IEEE,2016:647-651.
[4]Luo S,Zhang J.Accelerating harris algorithm with GPU for corner detection[C]//Seventh International Conference on Image and Graphics.IEEE Computer Society,2013:149-153.
[5]Khronos OpenCL Working Group.OpenCL-The open standard for parallel programming of heterogeneous systems[EB/OL].https://www.khronos.org/opencl,2015.
[6]Doxygen.CUDA-Accelerated Computer Vision[EB/OL].[2015-12-18].http://docs.opencv.org/3.1.0/d1/d1e/group__cuda.html.
[7]AMD Corporation.AMD APP SDK OpenCL optimization guide[S].August,2015.
[8]Gaster B,Howes L,Kaeli DR,et al.Heterogeneous computing with OpenCL:Revised OpenCL 1.2edition[M].Morgan Kaufmann Publishers Inc,2012:60-63.
[9]Du P,Weber R,Luszczek P,et al.From CUDA to OpenCL:Towards a performance-portable solution for multi-platform GPU programming[J].Parallel Computing,2012,38(8):391-407.
[10]Brodtkorb AR,Hagen TR,Stra ML.Graphics processing unit(GPU)programming strategies and trends in GPU computing[J].Journal of Parallel and Distributed Computing,2013,73(1):4-13.
[11]ZHANG Long.Research on parallel and performance optimization of canny algorithm based on OpenCL[D].Beijing:University of Chinese Academy of Sciences,2013(in Chinese).[张龙.基于OpenCL的Canny算法并行与性能优化研究[D].北京:中国科学院大学,2013.]
[12]Kalra A,Chhokar RL.A hybrid approach using Sobel and Canny operator for digital image edge detection[C]//International Conference on Micro-Electronics and Telecommunication Engineering.Ghaziabad:IEEE,2017:305-310.
[13]Ram P,Padmavathi S.Analysis of Harris corner detection for color images[C]//International Conference on Signal Processing, Communication, Power and Embedded System.Paralakhemundi:IEEE,2017:405-410.
[14]Bauer S,Leon FP.Using image pyramids for the acceleration of spectral unmixing based on nonnegative matrix factorization[C]//8th Workshop on Hyperspectral Image and Signal Processing:Evolution in Remote Sensing.Los Angeles:IEEE,2016:1-5.
[2]Jain A,Namdev A,Chawla M.Parallel edge detection by SOBEL algorithm using CUDA C[C]//IEEE Students’Conference on Electrical,Electronics and Computer Science.Bhopal:IEEE,2016:1-6.
[3]Huang Y,Bai Y,Li R,et al.Research of Canny edge detection algorithm on embedded CPU and GPU heterogeneous systems[C]//International Conference on Natural Computation,Fuzzy Systems and Knowledge Discovery.IEEE,2016:647-651.
[4]Luo S,Zhang J.Accelerating harris algorithm with GPU for corner detection[C]//Seventh International Conference on Image and Graphics.IEEE Computer Society,2013:149-153.
[5]Khronos OpenCL Working Group.OpenCL-The open standard for parallel programming of heterogeneous systems[EB/OL].https://www.khronos.org/opencl,2015.
[6]Doxygen.CUDA-Accelerated Computer Vision[EB/OL].[2015-12-18].http://docs.opencv.org/3.1.0/d1/d1e/group__cuda.html.
[7]AMD Corporation.AMD APP SDK OpenCL optimization guide[S].August,2015.
[8]Gaster B,Howes L,Kaeli DR,et al.Heterogeneous computing with OpenCL:Revised OpenCL 1.2edition[M].Morgan Kaufmann Publishers Inc,2012:60-63.
[9]Du P,Weber R,Luszczek P,et al.From CUDA to OpenCL:Towards a performance-portable solution for multi-platform GPU programming[J].Parallel Computing,2012,38(8):391-407.
[10]Brodtkorb AR,Hagen TR,Stra ML.Graphics processing unit(GPU)programming strategies and trends in GPU computing[J].Journal of Parallel and Distributed Computing,2013,73(1):4-13.
[11]ZHANG Long.Research on parallel and performance optimization of canny algorithm based on OpenCL[D].Beijing:University of Chinese Academy of Sciences,2013(in Chinese).[张龙.基于OpenCL的Canny算法并行与性能优化研究[D].北京:中国科学院大学,2013.]
[12]Kalra A,Chhokar RL.A hybrid approach using Sobel and Canny operator for digital image edge detection[C]//International Conference on Micro-Electronics and Telecommunication Engineering.Ghaziabad:IEEE,2017:305-310.
[13]Ram P,Padmavathi S.Analysis of Harris corner detection for color images[C]//International Conference on Signal Processing, Communication, Power and Embedded System.Paralakhemundi:IEEE,2017:405-410.
[14]Bauer S,Leon FP.Using image pyramids for the acceleration of spectral unmixing based on nonnegative matrix factorization[C]//8th Workshop on Hyperspectral Image and Signal Processing:Evolution in Remote Sensing.Los Angeles:IEEE,2016:1-5.