GPU-accelerated LDPC Decoding For Mobile Device And Implementation

Communication has become the most important part of human life nowadays,and new technologies are emerging.However,there are also some traditional techniques that need to be constantly improved to accommodate the current mobile communication demands.Channel code is very important to increase the reliability of communication system.Low Density Parity Check code is a kind of error correcting code that can be close to the Shannon limit and it is widely used in wireless communication nowadays.However,compared with other error correcting code,processing the LDPC code need massive computation,especially in the decoding stage.In wireless communication,the current LDPC decoder on mobile device is generally based on hardware implementation.Nonetheless,the flexibility of hardware platform is poor,and it is difficult to support LDPC codes with different code rates and different code length.Also,it is difficult to upgrade.First of all,this paper proposes a software LDPC decoder based on multithreading of Graphics Processing Unit(GPU)to support parallel multi-code word decoding.In order to improve the decoding speed of normalized minimum sum decoding algorithm,compression and segmentation of check matrix are stored in the GPU's private memory.The storage scheme of the confidence information is improved,thus saving the GPU memory resources.Secondly,aim at quasi-cyclic LDPC codes that more commonly used in wireless communication,a layered decoding algorithm based on GPU acceleration is designed to achieve faster convergence speed.The algorithm is designed to reduce unnecessary synchronization to achieve accelerated decoding.Finally,the decoding performance of LDPC code including decoding accuracy,decoding convergence speed,maximum throughput rate and decoding delay is tested in the actual mobile device environment.The experiment shows that LDPC decoding on the GPU of mobile device does not affect the accuracy of decoding,and GPU can speed up the decoding.In contrast,the layered decoding algorithm has faster convergence speed and greater maximum throughput,while the normalized minimum sum decoding algorithm has lower time delay.GPU for LDPC decoding on a mobile device can achieve 100 Mbps throughput as well as less than 1.6 millisecond delay,which is able to satisfy the work scenarios that is most strict with bandwidth and delay,such as voice chat.