Professional Documents
Culture Documents
ẋ ij ( t ) = –x ij ( t ) +∑ A i, j, k, l ( x k, l ( t ), y k, l ( t ) ) (2)
A Cellular Neural Network (CNN) is a large scale
C kl ∈ N ij
array of locally connected non-linear analog cells [1]-[2].
Since their initial disclosure in 1988, numerous applica- The { A i, j, k, l } are nonlinear functions of the state
tions of CNNs have been proposed. In this paper we use x i + k, j + l and output y i + k, j + l voltages of the cells in N ij ,
CNNs to implement large arithmetic arrays where continu-
including the cell itself.
ous speed/power trade-offs and lower slew rates (lower
The output equation is given by eqn. (3):
noise) are advantageous. Due to their local connectivity
the VLSI implementation of such networks is quite prom- 1
y ij ( t ) = f ( x ij ( t ) ) = --- ( x ij ( t ) + 1 – x ij ( t ) – 1 ) (3)
ising [3]-[4]. 2
CNNs are introduced in section 2. In section 3 we Note that y ij ( t ) ≤ 1 for all t ≥ 0 . The functions { A i, j, k, l }
first briefly introduce DBNS addition and explain the con-
are assumed to be space invariant, so that
ditions required to guarantee carry-free addition by the
A i, j, k, l = A k – i, l – j . For a r -neighborhood, the values of
application of two simple reduction rules. In section 4 we
propose methods for designing CNN arrays to implement A k – i, l – j are usually shown as a ( 2r + 1 ) × ( 2r + 1 ) matrix
DBNS addition. We describe the sets of simple conditions known as feedback template, A. (we are using a modified
that must be met in order to force transition of the state definition for the feedback template).
voltage of the cells by switching sourcing or sinking cur-
rents in the cell. These conditions ensure stability and syn- The input signal is loaded as the initial state voltages of the
chronous transitions of the output voltages of the cells cells x ij ( 0 ) and after transient time the result (output sig-
during the application of reduction rules and also prevent
participation of a cell in more than one reduction at the nal) is available from the output state variables y ij ( ∞ ).
same time. In section 5 we discuss implementation proce-
dures and present basic CMOS circuits of the DBNS cell. 3. Double-Based Number System (DBNS)
In section 6 we provide an illustrative example of DBNS
reduction to guarantee carry-free addition.
The double base number representation has the form
of eqn. (4), where the digits d i, j ∈ ( 0, 1 ) .
j j+1 j j+1
2 2 2 2
∑ d i, j 2 3
i j
x = (4)
i, j
...
with D,E,F and G (Fig. 4 b). j
arbitrary position (i, j)
i with three of its neighbors
...
...
... ... at positions (i, j – 1) ,
A B D E (i + 1, j – 1) and (i + 1, j) in
.. ... ... ... our CNN (Fig. 6)
δ C ϕ λ F G
...
We define the following
functions:
...
...
...
Group 2 conditions ticipate at position A (Fig.
4 a)) in reduction rule 1 and i
Considering the map of a DBNS number, each square
position E (Fig. 4 b)) in
(cell) has a specific value, and participation of any active ... ...
reduction rule 2 at the same
cell in more than one reduction at the same time leads to
time. (see Fig. 12).
incorrect results. Therefore in cases where two or more
reduction patterns have a cell in common, only one of
...
them is allowed. This requirement is forced using group 2
conditions. Participation of an active cell in more than one Fig. 12 Case 3c
reduction can occur in the following cases:
Case 1 As we see, choosing one reduction set over the other
The active cell at leads to different but equivalent representations of the
j
DBNS number (nonuniqueness property of the DBNS [7]).
...
position ( i, j ) can partici-
pate in two applications of However, in a CNN implementation of these five cases, we
i
reduction rule 1, in the left must select the appropriate design in terms of complexity
and speed of operation. In our design, minimum hardware
set shown by the broken ... ...
line or in the right set complexity and higher speed are achieved by selecting the
shown by the solid line. set shown by the solid line in each case using the group 2
conditions. By checking the output voltage of the cells at
Case 2 positions δ , ϕ and λ (Fig. 4 a) and b)) and comparing
...
Fig. 8 Case 1 them with threshold values we select the sets shown by the
The active cell at j solid lines over the sets shown by the broken lines (see Fig.
...
position ( i, j ) can partici- 8-12). The group 2 conditions comprise the following:
pate in two applications of
y ϕ ≥ 0.9, y δ ≥ 0.9, y λ < 0.7 (17)
reduction rule 2, in the
upper set shown by the i ... ... Now we define conditions required for switching Isnk
broken line or in the lower (Isrc) on. The following five sets ( A A , A B , A D , A E , A F )
set shown by the solid line.
of conditions are required to be satisfied for Isnk to be
...
Case 3b
Fig. 10 Case 3a H i, j ;w = H i, j – 1 ;s = 1
An active cell can j AB : & (19)
...
Fig. 11 Case 3b
H i, j ;w = H i, j ;sw = H i, j ;s = 1
AE : & (21) VDD
y i + 1, j – 2 < – 0.7
Vb2
M5
M8
A F : H i – 1, j + 1 ;w = H i – 1, j + 1 ;sw = H i – 1, j + 1 ;s = 1 (22) Vb1
y i, j + 1 ≥ 0.9
M1 2 M13
Reduction Rule 2
_ _ __
M10 M11
Vs r c Vsr c
H i – 1, j ;w = H i – 1, j ;sw = H i – 1, j ;s = 1
AG : & (24)
y i, j – 2 < – 0.7 VDD
M15
As we see, the dynamic of a cell is determined by its inter-
action with the other cells in the 5 × 5 neighborhood. M1 4 M1 6
Vb4
-1 9. References
1
[1] L. O. Chua and L. Yang, “Cellular neural networks: Theory
s21 s22 s23
& applications,” IEEE Trans. CAS, 35, pp. 1257-1290, 1988.
0 [2] T. Roska, and L. O. Chua, “Cellular neural networks with
nonlinear and delay-type template elements and nonuniform
grids,” Int. J. Circ. Th. and Appl., vol. 20, pp. 469-482, 1992.
-1 [3] J. E. Varrientos, E. Sanchez-Sinencio, and J. Ramirez-
Angulo, “A current -mode Cellular Neural Network Imple-
1 mentation,” IEEE Trans.CAS-II, vol. 40, No. 3, March 1993.
s31 s32 s33 [4] E. Sanchez-Sinencio, J. Ramirez-Angulo, B. Linares-Bar-
ranco, and A. Rodriguez-Vazquez, “Operational transcon-
0 ductance amplifier-based nonlinear function syntheses,”
IEEE Journal of Solid-State Circuits, 24, 6, Dec. 1989.
[5] L. O. Chua, M. Hassler, G. Moschytz, and J. Neirynck,
-1 “Autonomous Cellular Neural Networks: A unified paradigm
1 for pattern formation and active wave propagation,” IEEE
Transactions on Circuits and Systems-I, 42, 10, Oct.1995.
s41 s42 s43 [6] K. Jordar, “A simple approach to Modeling Cross-Talk in
Integrated Circuits”, IEEE Journal of Solid-State Circuits,
0 vol. 29, NO. 10, October 1994, pp. 1212-1219.
[7] V. Dimitrov, S. Sadeghi-Emamchaie, G.A. Jullien, and W.C.
Miller, ”A near Canonic Double-Based Number System
-1 (DBNS) with applications in Digital Signal Processing,”
20n 40n 20n 40n 20 40n Proc. SPIE, Adv. Sig.l Proc. VI, vol. 2846 pp. 14-25.
[8] V.S. Dimitrov, G.A. Jullien, and W.C. Miller, ”Theory and
Fig. 16 Cell state voltages Application for a Double-Base Number System,” Proceed-
ings of 13th symposium on Computer Arithmetic, Asilomar,
California 1997, pp. 44-51
7. Conclusions [9] S. Sadeghi-Emamchaie, G.A. Jullien, and W.C. Miller,”
Arithmetic Arrays using Cellular Neural Network” Proceed-
In this paper, we have introduced the concepts behind ings of 31th Asilomar conference on Signals, Systems and
a novel CNN-based processor for implementing digital Computer, Asilomar, California Nov. 1997
arithmetic arrays in a Double-Base number system. We
have presented sets of simple conditions that are required
to force the transition of the state voltage by switching