基于矢量DSP的并行化卷积算法

为了提场卷积算法在矢量!字信号处理器(DSP)上的执行效率,提出了一种高效的并行化卷积算法——基2并行短卷积(PSC R2)算法。该算法采用了基2短卷积运算结构,摆脱了传统并行化卷积算法的直接结构,从而有效降低了算法的循环次!。基于该算法结构,还提出了矢量DSP专用指令以匹配卷积的运算结构,保障算法执行效率。通过实际评估,证明了该算法在时间复杂度上仅为传统的内循环矢量化(VIL)算法的43%,为外循环矢量化(VOL)算法的55%,并且在存储空间开销上能够与传统算法基本持平。利用该算法,可以大幅降低移动通信和数字信号处理中的卷积、相关、滤波运算的时间复杂度。

A parallelized convolution algorithm for vector digital signal processors

To improve the efficiency of the convolution computation on a vector digital signal processor (DSP),the radix2 parallelized short convolution (PSC R2),a highly efficient parallelized algorithms was proposed.The PSC R2 algorithm uses a structure of radix-2 short convolution,not a direct structure of the conventional convolution,so that the number of algorithm cycle is effectively reduced.Furthermore,application specific DSP instructions were proposed to guarantee the high efficiency of the parallelized algorithm.It is proved by empirical analysis that the PSC R2 algorithm has the low temporal complexity,which accounts for only 43％ of the traditional Vectorising the Inner Loop (VIL) algorithm and 55％ of the traditional Vectorising the Outer Loop (VOL) algorithm;and has nearly the same memory consumption as the two traditional algorithms.In practical applications,the proposed PSC R2 algorithm could significantly reduce the temporal complexity in convolution,correlation and filtering operation in mobile communications and digital signal processing.