Effects of Finite Register Length in the Signal Processing Arithmetic Unit of the AN/UYK-17.
Abstract
Because of finite register length, several kinds of error are introduced in NRL's Signal Processing Arithmetic Unit (SPAU). The SPAU has a 16-bit word length and uses two's-complement, fixed-point, truncated arithmetic. A statistical model for arithmetic roundoff is formulated and used in discussing sources of error in multiplication and addition. Simulation proves a useful tool for studying quantization effects in the two fundamental digital signal processing algorithms, the fast Fourier transform (FFT) and recursive filtering. Simulation results indicate that conditional block scaling should be used, that scaling other than scaling the input should be used for the 4-pole filter to utilize more completely the SPAU's 90-dB dynamic range (6 dB per bit), and that rounding arithmetic should be used in the next version of the SPAU. The possibility of a floating-point machine should be investigated. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 12, 1974
- Accession Number
- AD0781880
Entities
People
- Judith N. Froscher
Organizations
- United States Naval Research Laboratory