Signal Flow Graphs Linear Time Invariant Discrete Time Systems can be made up from the elements { Storage, Scaling, Summation } • Storage: (Delay, Register) T or z -1 xk • Scaling: (Weight, Product, Multiplier xk-1 A yk xk or xk yk A 1 yk = A.xk Professor A G Constantinides Signal Flow Graphs • Summation: (Adder, Accumulator) • + X X+Y + Y • A linear system equation of the type considered so far, can be represented in terms of an interconnection of these elements • Conversely the system equation may be obtained from the interconnected components (structure). 2 Professor A G Constantinides Signal Flow Graphs • For example yk a1 yk 1 a2 yk 2 bxk xk b yk z-1 yk-1 a1 a2 3 yk-2 Professor A G Constantinides Signal Flow Graphs • A SFG structure indicates the way through which the operations are to be carried out in an implementation. • In a LTID system, a structure can be: i) computable : (All loops contain delays) ii) non-computable : (Some loops contain no delays) 4 Professor A G Constantinides Signal Flow Graphs • Transposition of SFG is the process of reversing the direction of flow on all transmission paths while keeping their transfer functions the same. • This entails: – Multipliers replaced by multipliers of same value – Adders replaced by branching points – Branching points replaced by adders • For a single-input / output SFG the transpose SFG has the same transfer function overall, as the original. 5 Professor A G Constantinides Structures • STRUCTURES: (The computational schemes for deriving the input / output relationships.) • For a given transfer function there are many realisation structures. • Each structure has different properties w.r.t. • i) Coefficient sensitivity • ii) Finite register computations 6 Professor A G Constantinides Signal Flow Graphs Direct form 1 : Consider the transfer function n i a . z i Y ( z) H ( z) i 0m X ( z ) 1 b . z i • So that • Set 7 i 1 i m n i Y ( z ). 1 bi .z X ( z ). ai .z i i 1 i 0 n W ( z ) X ( z ). a1.z i i 0 Professor A G Constantinides Signal Flow Graphs • For which a0 z-1 a1 z-1 z-1 a2 an + + • Moreover n delays + + W(z) m Y ( z ) W ( z ) bi .z i .Y ( z ) i 1 8 Professor A G Constantinides Signal Flow Graphs • For which W(z) Y(z) + + - z-1 b1 b2 z-1 z-1 b3 z-1 9 bm m delays Professor A G Constantinides Signal Flow Graphs • This figure and the previous one can be combined by cascading to produce overall structure. • Simple structure but NOT used extensively in practice because its performance degrades rapidly due to finite register computation effects 10 Professor A G Constantinides Signal Flow Graphs • Canonical form: Let H ( z ) H1 ( z ).H 2 ( z ) W ( z) 1 H1 ( z ) m X ( z) 1 b . z i • ie i 1 Y ( z) n H 2 ( z) ai .z i W ( z ) i 0 i m W ( z ) X ( z ) bi .z i .W ( z ) i 1 • and 11 n Y ( z ) ai .z i .W ( z ) i 0 Professor A G Constantinides Signal Flow Graphs • Hence SFG (n > m) a0 a1 a2 X(z) + + - - W(z) + + an + Y(z) + b1 b2 bm 12 Professor A G Constantinides Signal Flow Graphs • Direct form 2 : Reduction in effects due to finite register can be achieved by factoring H(z) and cascading structures corresponding to factors • In general H ( z ) H i ( z ) i with 1 2 • or 13 a0i a1i .z a2i .z Hi ( z) 1 b1i .z 1 b2i .z 2 1 a0i a1i .z Hi ( z) 1 b1i .z 1 Professor A G Constantinides Signal Flow Graphs k • Parallel form: Let H ( z ) g H i ( z ) i 1 • with Hi(z) as in cascade but a0i = 0 • With Transposition many more structures can be derived. Each will have different performance when implemented with finite precision 14 Professor A G Constantinides Signal Flow Graphs • Sensitivity: Consider the effect of changing a multiplier on the transfer function U(z) V(z) 2 1 X(z) 4 3 Y(z) Linear T-I Discrete System • Set 15 V ( z ) a. X ( z ) b.U ( z ) Y ( z ) c. X ( z ) d .U ( z ) • With constraint U ( z ) .V ( z )Professor A G Constantinides Signal Flow Graphs a • Hence V ( z ) .X ( z) 1 b. And Y ( z ) a c d . . G( z) X ( z) 1 b. thus G ( z ) da(1 b ) ad (b) (1 b ) 2 a d . 1 b 1 b 16 Professor A G Constantinides Signal Flow Graphs • Two-ports X1(z) Y1(z) 17 X2(z) Linear Systems T(z) S Y2(z) Professor A G Constantinides Signal Flow Graphs • Example: Complex Multiplier j y1 jy2 ( x1 jx2 )( j ) x1(n) y1(n) M x2(n) y2(n) M 18 Professor A G Constantinides Signal Flow Graphs • So that y (n) M x ( n) y1 (n) x1 (n) x2 (n) y2 (n) x1 (n) x2 (n) • Its SFD can be drawn as + x1(n) x2(n) 19 + - y1(n) + + + y2(n) Professor A G Constantinides Signal Flow Graphs • Special case 2 2 1 • We have a rotation of x(n) t o y (n) by an angle 1 tan • We can set cos 0 so that sin 0 and 0 20 • • • • This is the basis for designing i) Oscillators ii) Discrete Fourier Transforms (see later) iii) CORDIC operators in SONARProfessor A G Constantinides Signal Flow Graphs • Example: Oscillator • Consider y (n) M x(n) and externally impose the constraint x ( n) D So that I M • For oscillation det 21 D I M y ( n) y ( n) 0 D 0 Professor A G Constantinides Signal Flow Graphs • Set • Hence detI M z 1 D 0 0 1 z 1 z 1 z 1 D det 1 1 1 z z z 1 2 2 2 z z 1 z 1 2 22 1 2 1 2 Professor A G Constantinides Signal Flow Graphs • With 2 2 1 and cos 0T , 0 the oscillation frequency • Set x1 (n) cos 0nT then y1 (n) cos 0nT x2 (n) 23 and x1 (n) y1 (n 1) • We obtain x2 (n) sin0nT • Hence x1(n) and x2(n) correspond to two sinusoidal oscillations at 90 w.r.t. each other Professor A G Constantinides Signal Flow Graphs Alternative SFG with three real multipliers x1 (n) y1 (n) + + + + + x2 (n) 24 ( ) + y2 ( n ) Professor A G Constantinides