Architecture-dependent properties of analysis multirate filter banks

Analysis multirate filter banks (MRBs) are implemented with parallel, hierarchic, serial, and other architectures. The architectures vary widely in their properties. The authors generalize these architectures by defining a space of hybrid MRB architectures in which each member has different degrees of parallel, hierarchic, and serial processing. The authors calculate tradeoffs in processing time and hardware as a function of the architecture. On SISD (single instruction stream, single data stream) machines with direct form filters, they show that processing time decreases as the architecture becomes more hierarchic. Making the MRB more parallel only slightly lessens the time. With time-varying filters, making the MRB either more hierarchic or more parallel reduces the time. On hardwired machines, there is a parabolic tradeoff between processing time and hardware