支持任意系数长度和数据类型的FIR滤波器向量化方法

摘要

向量处理器的向量化算法映射是难点问题.提出一种支持任意系数长度和数据类型的FIR滤波器向量化方法,将(Finite Impulse Response)滤波器的卷积计算划分为系数长度步向量乘法和加法计算,每一步的向量乘法和加法计算在各个向量处理单元上并行执行,计算一个输出结果的所有乘法和加法计算都在同一向量处理单元上完成,每次循环能够同时完成向量处理单元数量个输出结果的计算.在向量处理器YHFT-Matrix的实验结果表明,该向量化FIR滤波器能够取得高效的计算性能和加速比.%The Vectorization of algorithm mapping for vector processors is a critical issue. This paper presents a vectorization of FIR filter supporting any different coefficients length and data types, in which the FIR filter computation is divided into N-step (length of coefficients)vector multiply and accumulate operations, vector multiply and accumulate operations are executed in parallel by all vector processing elements( VPEs), and all vector multiply and accumulate operations to calculate single output are completed at the same VPE, the number of VPEs outputs will be calculated at once loop. Experimental results on YHFT-Matrix show that the vectorization of FDR. filter achieves very efficient performance and speedup.