Improving Performance of High Precision Signal Processing Algorithms on Programmable DSPs

摘要

In applications, such as high quality audio, that need more than 16 bits of precision, the processing of signals on a 16-bit DSP requires double precision computation and is hence time consuming. Since in residue number system (RNS), the high precision data is decomposed to a lower precision for its processing, in this paper, we propose RNS for increasing the performance of such applications implemented on a single processor. We apply multirate architectures and also suggest some architectural extensions to the processor to further enhance the performance. The results show that performance improvement of more than 57% can be achieved with these implementations. We also show that the power dissipation can be significantly reduced with such implementations.