AN EFFICIENT VECTORIZATION OF FIR FILTER FOR VECTOR PROCESSOR

摘要

The Vectorization of algorithms mapping for vectorprocessor is a critical issues. An efficient vectorization ofFIR filter for vector processor is proposed, in which the FIRfilter computation is divided into N-step vectormultiplication, each vector multiplication is executed inparallel by sixteen vector processing elements, thecalculation of sixteen output is completed at once.Compared with existing methods, this method can fullyexploit the instruction level and data level parallelism ofvector processor, it can be applied to the FIR filter withdifferent length of coefficients, it is not limited to vectorprocessors whether to support the addition of reduction, andsupports 8-bit, 16-bit fixed-point real, fixed-point complex,32-bit floating-point real and complex data types.Experimental results show that the execution time forcalculating 1024-points with a 50-tap fixed-point real FIRfilter based on YHFT-Matrix is only 7.4 us, thevectorization of floating-point complex FIR filter achievesnearly 8x speedup over sequential algorithm ofTMS320C67x, the vectorization of fixed-point complexFIR filter achieves nearly 16x speedup over sequentialalgorithm of TMS320C64x.