基于YHFT-Matrix的FFT向量化设计与实现

发布时间：2019-04-02 04:54

【摘要】：随着LTE、4G无线通信、高清视频编解码、图像匹配、雷达信号处理等高性能计算应用的不断涌现，标量单核处理器难以满足这类应用的高密度实时计算需求，向量多核体系结构成为当前处理器设计的主流。但是，基于向量多核处理器的算法实现在并行编程、存储管理等方面面临巨大挑战，如何高效地开发向量多核处理器的多级并行性成为当前的研究热点。 YHFT-Matrix是国防科学技术大学自主研发的一款面向软件无线电的高性能向量多核处理器；而FFT/IFFT是LTE、4G无线通信技术中OFDM调制解调、MIMO信道估计等应用的核心算法，因此，面向YHFT-Matrix的向量多核体系结构特点，研究高效的FFT/IFFT向量化设计与实现方法具有重要的理论意义和应用价值。本文的主要工作包括：（1）针对单核YHFT-Matrix的向量数据访问、处理和混洗网络等体系结构特点，分别提出高效的基2、基3、基4和基5FFT的向量化方法，该方法利用FFT算法的内在并行性，能够充分挖掘YHFT-Matrix的指令级、数据级和多核多级并行性。实验结果表明，这些算法均取得了较高的计算性能和加速比，其中2K点基2FFT计算时间为2985拍，相对同频率下的TIC62xx，加速比为15.3；64K点基4FFT计算时间为91643拍，相对同频率下的TIC62xx，，加速比为14.48；（2）在上述的基础上，进一步提出一种实现混合基FFT的向量化方法，实验结果表明，1200点混合基FFT计算时间为1982拍，取得了较高的计算性能；（3）针对四核YHFT-Matrix的SDP硬件同步和Qlink大块数据传输的多核数据通信特点，提出一种实现四核并行的基2FFT的向量化方法，实验结果表明，64K点四核并行基2FFT的计算时间为46953拍，与YHFT-Matrix单核相比，加速比为2.58，取得了较高的性能加速比；（4）综合实现了一个包含FFT、位反序、信道估计、MIMO均衡和IFFT核心算法的OFDM接收端应用系统，实验结果表明，该系统取得了较高的计算性能，每个子帧的计算时间为234us，完全满足LTE所要求的1ms的时间需求。
[Abstract]:With the development of high-performance computing applications such as LTE,4G wireless communication, high-definition video codec, image matching, radar signal processing and so on, scalar single-core processors are difficult to meet the high-density real-time computing needs of such applications. Vector multi-core architecture has become the mainstream of current processor design. However, the implementation of algorithms based on vector multicore processors faces great challenges in parallel programming and memory management. How to efficiently develop multilevel parallelism of vector multicore processors has become a hot research topic. YHFT-Matrix is a high performance vector multicore processor for software radio developed by the National Defense University of Science and Technology. FFT/IFFT is the core algorithm of OFDM modulation and demodulation, MIMO channel estimation and so on in LTE,4G wireless communication technology. Therefore, the YHFT-Matrix-oriented vector multi-core architecture is characterized. The research on efficient FFT/IFFT vectorization design and implementation method has important theoretical significance and application value. The main work of this paper is as follows: (1) aiming at the architecture characteristics of single core YHFT-Matrix, such as vector data access, processing and shuffling network, the efficient vectorization methods of radix 2, 3, 4 and 5 FFT are proposed, respectively. This method makes use of the inherent parallelism of FFT algorithm, and can fully mine the instruction level, data level and multi-core multilevel parallelism of YHFT-Matrix. Experimental results show that these algorithms have high computational performance and speedup, in which the calculation time of 2K-point-based 2FFT is 2985 beats, and the relative TIC62xx, speedup at the same frequency is 15.3. The calculation time of 64K point-based 4FFT is 91643 beats, and the TIC62xx, acceleration ratio at the same frequency is 14.48. (2) on the basis of the above, a vectorization method for the implementation of mixed-base FFT is proposed. The experimental results show that the calculation time of 1200-point mixed-base FFT is 1982 beat, and high computational performance is obtained. (3) aiming at the characteristics of SDP hardware synchronization of quad-core YHFT-Matrix and multi-core data communication of Qlink bulk data transmission, a vectorization method for realizing four-core parallel radix-2 FFT is proposed. The experimental results show that: The calculation time of 64K point four-core parallel basis 2FFT is 46953 beat, compared with YHFT-Matrix single core, the speedup is 2.58, and the performance speedup is higher. (4) A OFDM receiver application system which includes FFT, bit inversion, channel estimation, MIMO equalization and IFFT core algorithm is implemented. The experimental results show that the system achieves high computational performance. The calculation time of each subframe is 234us, which fully meets the time requirement of 1ms required by LTE.
【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP332;TN929.5

【共引文献】