The Parallel Implementation Of LDPC-Codes Encoding And Stacked Autoencoder

As it is approaching to the physical limit of the integrated circuit chip,based on a comprehensive tradeoff between energy consumption and cost,the computing architecture has been developed in the direction of multi-core and many-core processors in recent years.To take full advantage of the computational resources of the modern computing architectures,algorithms should be specially designed for multi-core and many-core architectures.With the special program for applied research on super computation of the NSFC-Guangdong Joint Fund,this paper implemented parallel algorithms on the calculation of the minimum code weight of some LDPC codes and on stacked autoencoder-based saliency estimation on Tianhe-2.Parallel implementation of the LDPC codes encoding algorithm.As a channel coding standard of various advanced communication systems,LDPC codes are often used to correct errors occurring in transmission.By determining the minimum number of LDPC codes,researchers can design a better LDPC code structure.When calculating the number of minimum weight codes for a class of LDPC codes,the computational complexity increases exponentially along with the code length.In this paper,the algorithm is formulated as multiple-tree search with branching,which naturally provides the statistics of the minimum weight of any length LDPC code.In implementation,different types of computing cores are used on each subtree for parallel search.The final choice is the MIC and CPU collaborative computing model on Tianhe-2,which is used to obtain the results.Parallel implementation of the stacked autoencoder algorithm.Investigation to the computational model of visual saliency is helpful to improve our understanding to the human visual system.However,with the emergence of large-scale image database,visual saliency estimation based on stacked autoencoder could be very time-consuming with massive images.To implement parallel autoencoder,it is a good choice to use open source deep learning libraries.In this paper,we compared the advantages and disadvantages of libraries including Xeon-Caf Phi,DL4 J,Tensor Flow,and MXNet in for the algorithm to be implemented.Tensor Flow and MXNet are selected to parallel accelerate the stacked autoencoder.After testing on a local computer,the Tensor Flow-based implementation is twice faster than the original Matlab version.On Tianhe-2,tests show that the MXNet-based implemention is good enough on various performance indicators.This paper provides parallel implementations of the minimum code weight calculation of some LDPC codes and the stacked autoencoder.The work accelerates the research of the properties of LDPC codes and visual saliency estimation,at least to some extent.In addition,the two parallel algorithms also enlarged the number of scientific computation algorithms that can run on Tianhe-2.