Canonical Signed Digits (CSD)

• Canonical Signed Digits (CSD) is another commonly used

technique for simplifying multiplications in FIR filters

• Reduces switching-power by reducing number of ones in

multiplier

• Typically a 33% reduction in the number of ones (and hence

additions) is obtained using CSD

• Compared to CSD, DCM has 40% better power dissipation

• Modifications to DCM is possible for near multiplier-less filters

Source: Intel

Prof. Kaushik Roy

@ Purdue Univ.
 Coefficient Reordering
X(n-1) X(n-2)
X(n-1) X(n-2) C2
C2

C1 C3
C1 C3

X(n) X(n-3) X(n) X(n-3)

C4 C0 C4
C0

X(n-7) X(n-4)
X(n-4)
X(n-7)
C7 C5 C7 C5

C6 C6
X(n-6) X(n-5)
X(n-6) X(n-5)

• We represent this problem using a graph in which vertices represent

the coefficients and edges represent the resources required when the
differential coefficient corresponding to the edge is used in the
computation
• The optimal solution is the well-known problem of finding the
Source: Intel
Hamiltonian cycle of smallest weight of this graph (minimum spanning
tree), thus achieving even lower complexity
Prof. Kaushik Roy
@ Purdue Univ.
 Implementation of DCMI

Compute the Hamiltonian cycle in G

K ={k0,…,kM-1}=set of ordering of indices of the coefficients
(e.g. In DCM K={0,1,2,3,…,M-1}.Δcj=c1+j-cj)

 ci  c  (n) (  
 (c  c ) x(n  k )  P n k i k i1
)
Then k i1 ck i and Pk i ki k i1 k i1

Source: Intel

Prof. Kaushik Roy

@ Purdue Univ.
 Leakage Across Different Dies

Step 1 : Move the delay elements into branches

Source: Intel

Prof. Kaushik Roy

@ Purdue Univ.
 Leakage Across Different Dies

Step 2 : Move the delay elements out of the branches and connect
the appropriate partial product.

•Overhead = M-1 Add operations.

Source: Intel

Prof. Kaushik Roy

@ Purdue Univ.