摘要
为满足G(Gigabytes)级像素帧的实时性处理需求,针对信号处理系统中处理计算量大、实时性要求高的特点,剖析了解算过程内在的数据并行特性,深入研究了基于计算阵列的谱图解算数据并行算法。提出了一种基于MPP(Massively Parallel Processor)计算机SIMD PE阵列的FFT的数据并行计算实现方法。首先根据FFT架构中的数据交互一致性,给出了数据并行计算的表达式。提出一种基于PE标识,进行条件操作的SIMD PE阵列数据并行实现方法。该方法不但省去了并行处理中的数据寻址时间开销,而且使得数据并行操作更为规则、简洁,满足了阵列操作规则性强的处理要求,大幅度地提高了MPP计算机并行计算处理速度。该方案是一种简洁有效的PE自治问题解决方案,以更合理的方法和更高的效率实现了常规经典算法,在数据并行计算领域中,无疑具有重要的理论意义和应用价值,将在嵌入式信号处理中发挥愈来愈重要的作用。
In order to satisfy the real-time processing requirement of G-level pixel frame,considering the intensive and real time compu- ting requirement of signal processing in embedded signal system,the inner data parallelism of the calculation process is analyzed,and the data parallel algorithms of spectrogram calculation based on computing array is also researched. A data parallel computation method im- plemented on SIMD PE array of MPP （ Massively Parallel Processor ） computer for FFT transform is presented. Based on the data compu- ting consistency of FFT,the expression of data parallel computing is given firstly. Then a method of data parallel computing based on SIMD PE array to execute conditional operations by using of PE identifier is proposed, which not only omits the time cost of addressing, but makes the data parallel operation more regular and compact （ only in computation statements and move statements）. It meets the fea- tures of high regularity required by SIMD and greatly improves MPP computer processing speed,which is also a simple and effective PE autonomy solution,realizing conventional classic algorithms with more rational method and higher efficiency. It has important theoretical significance and application value in the area of data parallel undoubtedly, which will play a more and more important role in embedded signal processing.
出处
《计算机技术与发展》
2017年第10期91-95,共5页
Computer Technology and Development
基金
陕西省科技统筹创新工程计划(2015KTTSGY04-05)
