Abstract: This paper proposes an efficient finite precision block floating point (BFP) treatment to the fixed coefficient finite impulse response (FIR) digital filter. The treatment includes effective implementation of all the three forms of the conventional FIR filters, namely, direct form, cascaded and par- allel, and a roundoff error analysis of them in the BFP format. An effective block formatting algorithm together with an adaptive scaling factor is pro- posed to make the realizations more simple from hardware view point. To this end, a generic relation between the tap weight vector length and the input block length is deduced. The implementation scheme also emphasises on a simple block exponent update technique to prevent overflow even during the block to block transition phase. The roundoff noise is also investigated along the analogous lines, taking into consideration these implementational issues. The simulation results show that the BFP roundoff errors depend on the sig- nal level almost in the same way as floating point roundoff noise, resulting in approximately constant signal to noise ratio over a relatively large dynamic range.
Abstract: we propose a new normalized LMS (NLMS) algorithm, which gives satisfactory performance in certain applications in comaprison with con-ventional NLMS recursion. This new algorithm can be treated as a block based simplification of NLMS algorithm with significantly reduced number of multi¬ply and accumulate as well as division operations. It is also shown that such a recursion can be easily implemented in block floating point (BFP) arithmetic, treating the implementational issues much efficiently. In particular, the core challenges of a BFP realization to such adaptive filters are mainly considered in this regard. A global upper bound on the step size control parameter of the new algorithm due to BFP implementation is also proposed to prevent overflow in filtering as well as weight updating operations jointly.
Abstract: A special case of floating point data representation is block
floating point format where a block of operands are forced to have a joint
exponent term. This paper deals with the finite wordlength properties of
this data format. The theoretical errors associated with the error model for
block floating point quantization process is investigated with the help of error
distribution functions. A fast and easy approximation formula for calculating
signal-to-noise ratio in quantization to block floating point format is derived.
This representation is found to be a useful compromise between fixed point
and floating point format due to its acceptable numerical error properties over
a wide dynamic range.