Searcharxiv⌕ Search

arXiv subjects

Qinglin Wang

Publications and source records attributed to Qinglin Wang.

20 records · Page 2Linked to original sources

NUMA-aware FFT-based Convolution on ARMv8 Many-core CPUs

Convolutional Neural Networks (CNNs), one of the most representative algorithms of deep learning, are widely used in various artificial intelligence applications. Convolution operations often take most of the computational overhead of CNNs. The FFT-based algorithm can improve the efficiency of convolution by reducing its algorithm complexity, there are a lot of works about the high-performance implementation of FFT-based convolution on many-core CPUs. However, there is no optimization for the non-uniform memory access (NUMA) characteristics in many-core CPUs. In this paper, we present a NUMA-aware FFT-based convolution implementation on ARMv8 many-core CPUs with NUMA architectures. The implementation can reduce a number of remote memory access through the data reordering of FFT transformations and the three-level parallelization of the complex matrix multiplication. The experiment results on a ARMv8 many-core CPU with NUMA architectures demonstrate that our NUMA-aware implementation has much better performance than the state-of-the-art work in most cases.

cs.DC↗

Algorithm and Architecture for Path Metric Aided Bit-Flipping Decoding of Polar Codes

Polar codes attract more and more attention of researchers in recent years, since its capacity achieving property. However, their error-correction performance under successive cancellation (SC) decoding is inferior to other modern channel codes at short or moderate blocklengths. SC-Flip (SCF) decoding algorithm shows higher performance than SC decoding by identifying possibly erroneous decisions made in initial SC decoding and flipping them in the sequential decoding attempts. However, it performs not well when there are more than one erroneous decisions in a codeword. In this paper, we propose a path metric aided bit-flipping decoding algorithm to identify and correct more errors efficiently. In this algorithm, the bit-flipping list is generated based on both log likelihood ratio (LLR) based path metric and bit-flipping metric. The path metric is used to verify the effectiveness of bit-flipping. In order to reduce the decoding latency and computational complexity, its corresponding pipeline architecture is designed. By applying these decoding algorithms and pipeline architecture, an improvement on error-correction performance can be got up to 0.25dB compared with SCF decoding at the frame error rate of $10^{-4}$, with low average decoding latency.

eess.SP↗