Chapter 11

DSP Algorithm
Implementations
清大電機系林嘉文
[email protected]
03-5731152 11-1
Original PowerPoint slides prepared by S. K. Mitra
© The McGraw-Hill Companies, Inc., 2007
 Matrix Representation of Digital
Filter Structures
• Consider
Co s de

• This structure is described by the set of equations

w1[[n]] = x[n]
[ ] −α w5[[n]]
w2[n] = w1[n] − δ w3[n]
w3[n] = w2 [n−1]
w4[n] = w3[n] + ε w2[n]
w5[n] = w4 [n−1]
y[n] = β w1[n] + γ w5[n]
Original PowerPoint slides prepared by S. K. Mitra 11-2
© The McGraw-Hill Companies, Inc., 2007
 Matrix Representation of Digital
Filter Structures
• Thiss o ordered
de ed set o of equat
equations
o s is
s sa
said
d to be noncomputable
o co putab e
because some variables on the left side are computed
before some right-side variables are made available
• For example, computation of w1[n] in the 1st step requires
the knowledge of w5[n] which is computed in the 5th step
• Likewise,
Likewise computation of w2[n] in the 2nd step requires the
knowledge of w3[n] that is computed in the 3rd step
• Supposepp we reorder the equations q
w3[n] = w2 [n−1]
w5[n] = w4 [n−1]
w1[n] = x[n] −α w5[n]
w2[n] = w1[n] − δ w3[n]
w4[n] = w3[n] + ε w2[n]
Original PowerPoint slides prepared by S.y[n] = β w1[n] + γ w5[n]
K. Mitra 11-3
© The McGraw-Hill Companies, Inc., 2007
 Matrix Representation
• Thiss new
e o ordered
de ed set oof equat
equations
o s is s co
computable
putab e
• In most practical applications, equations describing a digital
filter structure can be put into a computable order by
inspection
• A simple way to examine the computability of equations
describing a digital filter structure is by writing the equations
in a matrix form
• A matrix representation
p of the first ordered set of equations:
q

Original PowerPoint slides prepared by S. K. Mitra 11-4

© The McGraw-Hill Companies, Inc., 2007
 Matrix Representation
• In co
compact
pact form
o
y[n] = x[n] + F y[n] + G y[n−1]
where
y[n] = [w1[n] w2[n] w3[n] w4[n] w5[n] y[n]]T
x[n] = [x[n] 0 0 0 0 0]T

• If a diagonal element of F is nonzero

nonzero, then computation of
present value of the corresponding variable requires the
knowledgeg of its p
present value implying
py gp presence of a
delay-free loop 11-5
Original PowerPoint slides prepared by S. K. Mitra
© The McGraw-Hill Companies, Inc., 2007
 Matrix Representation
• Anyy nonzero
o eoe entries
t es in tthe
e sa
samee row
o abo
above e tthe
e main
a
diagonal of F imply that the computation of present value of
the corresponding variable requires present values of other
variables not yet computed, making the set of equations
noncomputable
• Hence,
Hence for computability all elements of F matrix on the
diagonal and above diagonal must be zeros
• In the F matrix for the first ordered set of equations,
q ,
diagonal elements are all zeros, indicating absence of
delay-free loops
• However, there are nonzero entries above the diagonal in
the first and second rows of F indicating that the set of
equations are not in proper order for computation
Original PowerPoint slides prepared by S. K. Mitra 11-6
© The McGraw-Hill Companies, Inc., 2007
 Matrix Representation
• The
e F matrix
at foro tthe
e seco
second
doordered
de ed set o
of equat
equations
o s is
s

which is seen to satisfyy the computability

p y condition

Original PowerPoint slides prepared by S. K. Mitra 11-7

© The McGraw-Hill Companies, Inc., 2007
 Precedence Graph
• The precedence graph can be used to test the
computability
t bilit off a di
digital
it l filt
filter structure
t t and
d tto d
develop
l ththe
proper ordering sequence for a set of equations describing
a computable structure
• It is developed from the signal-flow graph description of
g
the digital filter structure in which independent
p and
dependent signal variables are represented by nodes, and
the multiplier and delay branches are represented by
di t d branches
directed b h
• The directed branch has an attached symbol denoting the
branch gain or transmittance
• For a multiplier branch, the branch gain is the multiplier
coefficient value
• For a delay branch, the branch gain is simply z−1 11-8
Original PowerPoint slides prepared by S. K. Mitra
© The McGraw-Hill Companies, Inc., 2007
 Precedence Graph
• The signal-flow graph representation of

is shown below

Original PowerPoint slides prepared by S. K. Mitra 11-9

© The McGraw-Hill Companies, Inc., 2007
 Precedence Graph
• A reduced signal-flow graph is then developed by
removing
i ththe d
delay
l b branches
h andd allll b
branches
h going
i outt off
the input node
• The reduced signal-flow graph of the example digital filter
structure is shown below

Original PowerPoint slides prepared by S. K. Mitra 11-10

© The McGraw-Hill Companies, Inc., 2007
 Precedence Graph
• The nodes in the reduced signal-flow graph are grouped as
f ll
follows:
• All nodes with only outgoing branches are grouped into one
set labeled {N1}
• Next, the set {N2} is formed containing nodes coming in
onlyy from one or more nodes in the set {N
{ 1} and have
outgoing branches to the other nodes
• Then, form the set {N3} containing nodes that have
b
branches
h comingi iin onlyl ffrom one or more nodes
d iin th
the
sets {N1} and {N2}, and have outgoing branches to other
nodes
• Continue the process until there is a set of nodes {Nf}
containingg only
y incoming g branches
• The rearranged graph is called a precedence graph 11-11
Original PowerPoint slides prepared by S. K. Mitra
© The McGraw-Hill Companies, Inc., 2007
 Precedence Graph
• Since signal variables belonging to {N1} do not depend on
th presentt values
the l off other
th signal
i l variables,
i bl th
these variables
i bl
should be computed first
• Next,
Next signal variables belonging to {N2} can be computed
since they depend on the present values of signal variables
contained in {{N1} that have alreadyy been computed p
• This is followed by the computation of signal variables in
{N3}, {N4}, etc.
• Finally,
Fi ll iin ththe llastt step
t th the signal
i l variables
i bl iin {Nf} are
computed
• This process of sequential computation ensures the
development of a valid computational algorithm
• If there is no final set {{Nf} containing g only
y incomingg
branches, the digital filter structure is noncomputable 11-12
Original PowerPoint slides prepared by S. K. Mitra
© The McGraw-Hill Companies, Inc., 2007
 Precedence Graph

• For the example precedence graph, pertinent groupings of

node
d variables
i bl are:
{N1} = {w3[n], w5[n]}
{N2} = {w1[n]}
{N3} = {w2[n]}
{{N4} = {w
{ 4[[n],
] y[
y[n]}
]}
Original PowerPoint slides prepared by S. K. Mitra 11-13
© The McGraw-Hill Companies, Inc., 2007
 Precedence Graph
• Precedence graph redrawn according to the above
groupings
i iis as shown
h b
below
l

• Since the final node set {N4} has only incoming branches
branches,
the structure is computable

Original PowerPoint slides prepared by S. K. Mitra 11-14

© The McGraw-Hill Companies, Inc., 2007
 Structure Verification
• A simple method to verify that the structure developed is
i d d characterized
indeed h t i db by the
th prescribed
ib d transfer
t f function
f ti
H(z)
• Consider for simplicity a causal 3rd order IIR transfer
function

• If {h[n]} denotes its impulse response, then

• Note P(z) = H(z)D(z),

H(z)D(z) which is equivalent to

Original PowerPoint slides prepared by S. K. Mitra 11-15

© The McGraw-Hill Companies, Inc., 2007
 Structure Verification
• Evaluate above convolution sum for 0 ≤ n ≤ 6:
p0 = h[0]
p1 = h[1] + h[0]d1
p2 = h[2] + h[1]d1 + h[0]d2
p3 = h[3] + h[2]d1 + h[1]d2 + h[0]d3
0 = h[4]
[ ] + h[3]d
[ ] 1 + h[2]d
[ ] 2 + h[1]d
[ ] 3
0 = h[5] + h[4]d1 + h[3]d2 + h[2]d3
0 = h[6] + h[5]d1 + h[4]d2 + h[3]d3
• In
I matrix
t i form
f we gett

Original PowerPoint slides prepared by S. K. Mitra 11-16

© The McGraw-Hill Companies, Inc., 2007
 Structure Verification
• In partitioned form above matrix equation can be written as

• Where

• Solving second equation we get

• Substituting above in the first equation we get

• In the case of an N-th order IIR filter, the coefficients of its

transfer function can be determined from the first 2N+1
impulse response samples
Original PowerPoint slides prepared by S. K. Mitra 11-17
© The McGraw-Hill Companies, Inc., 2007
 Structure Verification
• Example: Consider the causal transfer function

• Here
h[0] = 2, h[1] = 4, h[2] = −5, h[3] = −3, h[4] =13
• Hence

• Solving the above we get

Original PowerPoint slides prepared by S. K. Mitra 11-18

© The McGraw-Hill Companies, Inc., 2007
 Four Classes of Fourier Transforms
Continuous in Time Discrete in time
time-Periodic
Periodic in frequency

FT DTFT; T0
f(k) F(ejωT0)
ntinuous in

t k ω
2π
frequency

T0
Con

1 ∞ T π

∫ ∫π F ( e jωT0 )e jkωT0 dω
jω t
f (t ) = F ( jω ) e dω f (k ) = 0 T0

2π −∞ 2π −
T0
∞
∫
∞
F (ω ) = f ( t ) e − j ω t dt
−∞ F (e jωT0
)= ∑ f ( k )e
k = −∞
− jkωT0

Fourier transform Discrete time Fourier Transform

Original PowerPoint slides prepared by S. K. Mitra 11-19
© The McGraw-Hill Companies, Inc., 2007
 Four Classes of Fourier Transforms
Continuous in Time Discrete in time-Periodic in frequency
q y
F(n)
dic in

f(t) F(n)
F(k)
ete in frequency – period

t nω0 k n

2π
ω0 DFT:N
FSiω0
∞ N −1 2π
1
∑ ∑ F ( n )e
j kn
jnω 0 t
f (t ) = F ( n ) e f (k ) = N

n = −∞ N n =0

ω π N −1 2π
Discre

−j
∫ F ( n ) = ∑ f ( k )e
ω0 − jn ω 0 t kn
F (n ) = 0 f (t )e dt N
time

2π −π
ω0 k =0

Fourier series Discrete Fourier Transform

Original PowerPoint slides prepared by S. K. Mitra 11-20
© The McGraw-Hill Companies, Inc., 2007
 Direct Computation of DFT
For a complex-valued sequence of N C
C DFT SUBROUTINE
points
i the
h DFT may bbe expressed
d as C ISEL = 0 : DFT
C ISEL = 1 : INVERSE DFT
N −1
⎡ 2πkn 2πkn ⎤
X R (k ) = ∑ ⎢ xR (n) cos
C
+ xi (n) sin
N ⎥⎦
SUBROUTINE DFT(N, XR, XI, XFR, XFI, ISEL)
n =0 ⎣ N DIMENSION XR(N), XI(N), XFR(N), XFI(N)
N −1
⎡ 2πkn 2πkn ⎤ WN = 6.2831853 / FLOAT(N)
X I (k ) = −∑ ⎢ xR (n) sin − xi (n) cos
N ⎥⎦
IF (ISEL.EQ.1) WN = - WN
n =0 ⎣ N DO 20 K = 1, N
XFR(K) = 0,
• The direct computation XFI(K) = 0,
0
KM1 = K – 1
requires: DO 20 I = 1, N
IM1 = I – 1
• 2N2 evaluations of trigonometric ARG = WN * KM1 * IM1
functions. C = COS(ARG)
S = SIN(ARG)
• 4N2 real multiplications. XFR(K) = XFR(K) + XR(I)*C + XI(I)*S
XFI(K) = XFI(K) – XR(I)*S + XI(I)*C
• 4N(N –11 ) real additions
additions. 10
0 CONTINUE
CO U
IF (ISEL – 1) 20, 30, 20
• A number of indexing and 30
20
XFR(K) = XFR(K) / FLOAT(N)
XFI(K) = XFI(K) / FLOAT(N)
addressing operations. CONTINUE
RETURN
END
Original PowerPoint slides prepared by S. K. Mitra 11-21
© The McGraw-Hill Companies, Inc., 2007
 Goertzel’s
Goertzel s Algorithm
• A recursive DFT computation scheme that makes use of
the identity

obtained using the periodicity of

• Using this identity we can write

Original PowerPoint slides prepared by S. K. Mitra 11-22

© The McGraw-Hill Companies, Inc., 2007
 Goertzel’s
Goertzel s Algorithm
• Define
• Note: yk[n] is the direct convolution of the causal sequence

with a causal sequence

• Observe
Ob X[k] = yk[n]|
[ ]|n=N
• z-transform of yields

Original PowerPoint slides prepared by S. K. Mitra 11-23

© The McGraw-Hill Companies, Inc., 2007
 Goertzel’s
Goertzel s Algorithm
• Thus, yk[n] is the output of an initially relaxed LTI digital
filter Hk[z] with an input xe[n] and, when n = N, yk[n] = X[k]
• Structural interpretation of the algorithm

• Thus a recursive DFT computation scheme is

with yk[−1] = 0 and xe[N] = 0

Original PowerPoint slides prepared by S. K. Mitra 11-24

© The McGraw-Hill Companies, Inc., 2007
 Goertzel’s
Goertzel s Algorithm
• Since a complex multiplication needs 4 real multiplications
and 2 real additions, computation of each new value of yk[n]
requires 4 real multiplications and 4 real additions
• Thus computation of X[k] = yk[N] involves 4N real
multiplications and 4N real additions
• Computation
C t ti off allll N DFT samples l requires
i 4N2 reall
multiplications and 4N2 real additions
• Direct computation of all N samples of {X[k]} requires N2
complex multiplications and N(N-1) complex additions
• Equivalently,
Equivalently direct computation of all N samples of {X[k]}
requires 4N2 real multiplications and N(4N-2) real additions
• Thus, GoertzelGoertzel’s s algorithm requires 2N more real additions
than the direct DFT computation
Original PowerPoint slides prepared by S. K. Mitra 11-25
© The McGraw-Hill Companies, Inc., 2007
 Goertzel’s
Goertzel s Algorithm
• The algorithm can be made computationally more efficient
by observing that Hk[z] can be rewritten as

resulting in a second-order realization

Original PowerPoint slides prepared by S. K. Mitra 11-26

© The McGraw-Hill Companies, Inc., 2007
 Goertzel’s
Goertzel s Algorithm
• DFT computation equations are now

• Computation of each sample of involves only 2 real

multiplications and 4 real additions
• Complex multiplication by needs to be performed only
once at n = N
• Thus, computation of one sample of X[k] requires (2N+4)
real multiplications and (4N+4) real additions
• Computation of all N DFT samples requires 2N(N+2) real
multiplications and 4N(N+1) real additions

Original PowerPoint slides prepared by S. K. Mitra 11-27

© The McGraw-Hill Companies, Inc., 2007
 Goertzel’s
Goertzel s Algorithm
• In realizing HN−k[z], the multiplier in the feedback path is
2cos(2π(N − k)/N) = 2cos(2πk/N)
which is same as that in the realization of Hk[[z]]
⇒ νN−k[n] = νk[n], i.e., the intermediate variables computed
to determine X[k] can again be used to determine X[N−k]
• Only difference between the two structures is the feed-
forward multiplier which is now that is the complex
conjugate
j t off
• Thus, computation of X[k] and X[N−k] require 2(N+4) real
multiplications and 4(N+2) real additions
• Computation of all N DFT samples require approximately N2
real multiplications and approximately 2N2 real additions
Original PowerPoint slides prepared by S. K. Mitra 11-28
© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• Consider a sequence x[n] of length N = 2μ
• Using a 2-band polyphase decomposition we can express
its z-transform as
X( ) = X0(z
X(z) ( 2) + z−11X1(z
( 2)
where

• Evaluating on the unit circle at N equally spaced points

we arrive at the N-point
p DFT of x[n]:
[ ]

where X0[[k]] and X1[[k]] are the (N/2)-point

( )p DFTs of x0[[n]] and
x1[n] 11-29
Original PowerPoint slides prepared by S. K. Mitra
© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• That is

• Block diagram interpretation:

Original PowerPoint slides prepared by S. K. Mitra 11-30

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• Flow-graph representation

Original PowerPoint slides prepared by S. K. Mitra 11-31

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• Direct computation of the N-point DFT requires N2 complex
multiplications
lti li ti and d N2 − N ≈ N2 complex
l additions
dditi
• The modified scheme requires the computation of two
(N/2) point DFTs that are then combined with
(N/2)-point ith N complex
comple
multiplications and N complex additions resulting in a total
of (N2/2) + N complex multiplications and approximately
(N2/2) + N complex additions
• For N ≥ 3,, ((N2/2)) + N < N2

Original PowerPoint slides prepared by S. K. Mitra 11-32

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• Continuing the process we can express and as a weighted
combination
bi ti off ttwo (N/4)
(N/4)-point
i t DFT
DFTs
• For example, we can write

where X00[k] and X01[k] are the (N/4)-point DFTs of the

(N/4) l
(N/4)-length
th sequences x00[n]
[ ] = x0 [2n]
[2 ] and
d x01[n]
[ ] = x0 [2n+1]
[2 1]
• Likewise, we can express

where X10[k] and X11[k] are the (N/4)-point DFTs of the

(N/4)-length sequences x10[n] = x1 [2n] and x11[n] = x1 [2n+1]

Original PowerPoint slides prepared by S. K. Mitra 11-33

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• Block-diagram representation of the two-stage algorithm

Original PowerPoint slides prepared by S. K. Mitra 11-34

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• Flow-graph representation

Original PowerPoint slides prepared by S. K. Mitra 11-35

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• In the flow-graph shown N = 8
• Hence, the (N/4)-point DFT here is a 2-point DFT and no
further decomposition is possible
• The four 2-point DFTs, Xij[k], i, j = 0,1 can be easily
computed
• For
F example l

• C
Corresponding
di flflow-graph
h off th
the 2-point
2 i t DFT is
i shown
h
below obtained using the identity

Original PowerPoint slides prepared by S. K. Mitra 11-36

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• Complete flow-graph of the 8-point DFT is shown below

Original PowerPoint slides prepared by S. K. Mitra 11-37

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• The flow-graph consists of 3 stages
• First stage computes the four 2-point DFTs
• Second stage computes the two 4-point DFTs
• Last stage computes the desired 8-point DFT
• The number of complex multiplications and additions at
each stage is equal to 8, the size of the DFT
• Total number of complex multiplications and additions to
compute t allll 8 DFT samples
l iis equall tto 8 + 8 + 8 = 24 = 8
8×3
3
• In the general case when N = 2μ, the number of stages for
computing the N-point
N point DFT in the algorithm will be μ = log2N
• Total number of complex multiplications and additions to
compute all N DFT samples is N(log2N)
Original PowerPoint slides prepared by S. K. Mitra 11-38
© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• Note, and
• Besides, the coefficients have the symmetry property of

• These properties can be exploited to reduce the

computational complexity further
• In the following basic module two output variables are
generated by a weighted combination of two input variables

• The basic computational module is called a butterfly

computation
t ti
Original PowerPoint slides prepared by S. K. Mitra 11-39
© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• The input-output relations of the basic module are:

• Substituting in the second equation given

above we get

• The modified butterfly computation requires only one

complex multiplication as indicated below

• Use of the above modified butterfly computation module

reduces the total number of complex multiplications by 50%
Original PowerPoint slides prepared by S. K. Mitra 11-40
© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• New flow-graph using the modified butterfly computational
module
d l ffor N = 8

Original PowerPoint slides prepared by S. K. Mitra 11-41

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• Computational complexity can be reduced further by
avoiding
idi multiplications
lti li ti b
by , , , and
d

• The DFT computation algorithm described here also is

efficient with regard to memory requirements
• Note:
N t E Eachh stage
t employs
l th
the same b butterfly
tt fl computation
t ti tto
compute Ψr+1[α] and Ψr+1[β] from Ψr[α] and Ψr[β]
• At the end of computation at any stagestage, output variables
Ψr+1[m] can be stored in the same registers previously
occupied by the corresponding input variables Ψr[m]
• This type of memory location sharing is called in-place
computation
p resulting
g in significant
g savings
g in overall
memory requirements 11-42
Original PowerPoint slides prepared by S. K. Mitra
© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• In the DFT computation scheme outlined, the DFT samples
X[k] appear att the
th output
t t in
i a sequential
ti l order
d while
hil th
the
input samples x[n] appear in a different order
• Thus,
Th s a sequentially
seq entiall ordered inp
inputt x[n]
[n] must
m st be reordered
appropriately before the fast algorithm described by this
structure can be implemented

Original PowerPoint slides prepared by S. K. Mitra 11-43

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• To understand the input reordering scheme, represent the
arguments t off input
i t samples
l x[n]
[ ] and
d their
th i sequentially
ti ll
ordered new representations Ψ1[m] in binary forms with the
relations between the arguments m and n as follows:

• Th
Thus, if (b2b1b0) representst the
th index
i d n off x[n],
[ ] then
th the
th
sample x[b2b1b0] appears at the location m = b0b1b2 as
Ψ1[b0b1b2] before the DFT computation is started
• i.e., the location of Ψ1[m] is in bit-reversed order from that
of x[n]
• Alternative forms can be obtained by reordering the
p
computations such as input
p in normal order and outputp in bit-
reversed order, and both input and output in normal order 11-44
Original PowerPoint slides prepared by S. K. Mitra
© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• The fast algorithm described assumes that the length of x[n]
i a power off 2
is
• If it is not, the length can be extended by zero-padding and
make the length a po powerer of 2
• Even after zero-padding, the DFT computation based on the
fast algorithm may be computationally more efficient than a
direct DFT computation of the original shorter sequence
• The fast DFT computation schemes described are called
decimation-in-time (DIT) fast Fourier transform (FFT)
algorithms as input x[n] is first decimated to form a set of
subsequences before the DFT is computed

Original PowerPoint slides prepared by S. K. Mitra 11-45

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
Complex Complex
For N = 2μ, this decimation can be N
multiplications in multiplications in
Direct Computation FFT algorithm,
performed μ = log2N times. N2 (N/2)log2N
4 16 4

The total number of 8 64 12

16 256 32
/ ) ⋅ log2N
Complex multiplications : ((N/2)
C
32 1,024 80
Complex additions : N ⋅ log2N
64 4,096 192

128 16,384 448

256 65,535 1,024

512 262,144 2,304

1,024 1,048,576 5,120

Original PowerPoint slides prepared by S. K. Mitra 11-46

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• For example, the relation between x[n] and its even and odd
parts,
t x0[n]
[ ] andd x1[n],
[ ] generated
t d by
b the
th fifirstt stage
t off the
th DIT
algorithm is given by

• Likewise, the relation between x[n] and the sequences

x00[n], x01[n], x10[n], and x11[n], generated by the two-stage
decomposition of the DIT algorithm is given by

Original PowerPoint slides prepared by S. K. Mitra 11-47

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Time FFT Algorithm
Decimation-in-Time
• The subsequences x00[n], x01[n], x10[n], and x11[n] can be
generated
t d di
directly
tl bby a ffactor-of-4
t f4d decimation
i ti process
leading to a single-stage decomposition as shown below:

Original PowerPoint slides prepared by S. K. Mitra 11-48

© The McGraw-Hill Companies, Inc., 2007
 Radix R FFT Algorithm
Radix-R
• Radix-R FFT algorithm - A each stage the decimation is by
a factor
f t off R
• Depending on N, various combinations of decompositions of
X[k] can be usedsed to develop
de elop different ttypes
pes of DIT FFT
algorithms
• If the scheme uses a mixture of decimations by different
factors, it is called a mixed radix FFT algorithm
• For N which is a composite number expressible in the form
of a product of integers:
N = r1‫ ڄ‬r2 ‫ ڄ … ڄ‬r ν
the total number of complex multiplications (additions) in a
DIT FFT algorithm
g based on a ν -stage g decomposition
p is
given by 11-49
Original PowerPoint slides prepared by S. K. Mitra
© The McGraw-Hill Companies, Inc., 2007
 Decimation in Frequency FFT Algorithm
Decimation-in-Frequency
• For k even

• For k odd

• We can write

where

Original PowerPoint slides prepared by S. K. Mitra 11-50

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Frequency FFT Algorithm
Decimation-in-Frequency
• Thus X[2l] and X[2l+1] are the (N/2)-point DFTs of the
l
length-(N/2)
th (N/2) sequences x0[n][ ] and
d x1[n],
[ ]
• The flow-graph of the first-stage of the DFT:

• The output DFT samples appear in a decimated form with

the even-indexed samples appearing as the output of one
(N/2)-point
(N/2) point DFT and the odd-indexed odd indexed samples appearing as
the output of the other (N/2)-point DFT
Original PowerPoint slides prepared by S. K. Mitra 11-51
© The McGraw-Hill Companies, Inc., 2007
 Decimation in Frequency FFT Algorithm
Decimation-in-Frequency
• We next express the even- and odd-indexed samples of
each
h one off the
th two
t (N/2)-point
(N/2) i t DFTs
DFT as a sum off two
t (N/4)-
(N/4)
point DFTs
• The process is continued
contin ed until
ntil the smallest DFTs are 22-point
point
DFTs
• The flow-graph
flow graph of the decimation
decimation-in-frequency
in frequency (DIF) FFT
computation scheme for N = 8

Original PowerPoint slides prepared by S. K. Mitra 11-52

© The McGraw-Hill Companies, Inc., 2007
 Decimation in Frequency FFT Algorithm
Decimation-in-Frequency
• Computational complexity of the radix-2 DIF FFT algorithm
i same as th
is thatt off the
th DIT FFT algorithm
l ith
• Various forms of DIF FFT algorithm can similarly be
de eloped
developed
• The DIT and DIF FFT algorithms described here are often
referred to as the Cooley-Tukey
Cooley Tukey FFT algorithms

Original PowerPoint slides prepared by S. K. Mitra 11-53

© The McGraw-Hill Companies, Inc., 2007
 Inverse DFT Computation
• An FFT algorithm for computing the DFT samples can also
b used
be d tto calculate
l l t efficiently
ffi i tl th
the iinverse DFT (IDFT)
• Consider a length-N sequence x[n] with an N-point DFT X[k]
• Recall

• Multiplying both sides by N and taking the complex

conjugate we get

• Right-hand side off above is the N-point DFT off a sequence

X*[k]

Original PowerPoint slides prepared by S. K. Mitra 11-54

© The McGraw-Hill Companies, Inc., 2007
 Inverse DFT Computation
• The desired IDFT x[n] is then obtained as

• Inverse DFT computation is shown below:

Original PowerPoint slides prepared by S. K. Mitra 11-55

© The McGraw-Hill Companies, Inc., 2007
 Cooley Tukey FFT Algorithms
Cooley-Tukey
• Consider a length-N sequence x[n] with an N-point DFT X[k]
where
h N = N1N2
• Represent the indices n and k as

• Using these index mappings we can write

as

Original PowerPoint slides prepared by S. K. Mitra 11-56

© The McGraw-Hill Companies, Inc., 2007
 Cooley Tukey FFT Algorithms
Cooley-Tukey
• Since , , and , we
have

where 0 ≤ k1 ≤ N1 −1 and 0 ≤ k2 ≤ N2 −1
• The effect of the index mapping is to map the 1-D sequence
x[n] into a 2-D sequence that can be represented as a 2-D
array with n1 specifying the rows and n2 specifying the
columns of the array
• Inner parentheses of the last equation is seen to be the set
of N1-point DFTs of the N2-columns:

Original PowerPoint slides prepared by S. K. Mitra 11-57

© The McGraw-Hill Companies, Inc., 2007
 Cooley Tukey FFT Algorithms
Cooley-Tukey
• Note: The column DFTs can be done in place
• Next, these row DFTs are multiplied in place by the twiddle
factors yielding

• Finally,
y, the outer sum is the set of N2-point
p DFTs of the
columns of the array:

• The row DFTs, X[k1+N1k1], can again

g be computed in place
• The input x[n] is entered into an array according to the index
map:

Original PowerPoint slides prepared by S. K. Mitra 11-58

© The McGraw-Hill Companies, Inc., 2007
 Cooley Tukey FFT Algorithms
Cooley-Tukey
• Likewise, the output DFT samples X[k] need to extracted
f
from the
th array according
di tto th
the iindex
d map:

• Example: Let N = 8. Choose N1 = 2 and N2 = 4, then

for 0 ≤ k1 ≤ 1 and 0 ≤ k2 ≤ 3
• 2-D array representation of the input is

Original PowerPoint slides prepared by S. K. Mitra 11-59

© The McGraw-Hill Companies, Inc., 2007
 Cooley Tukey FFT Algorithms
Cooley-Tukey
• The column DFTs are 2-point DFTs given by

• These DFTs require

q no multiplications
p
• 2-D array of row transforms is

• After multiplying by the twiddle factors array becomes

Original PowerPoint slides prepared by S. K. Mitra 11-60

© The McGraw-Hill Companies, Inc., 2007
 Cooley Tukey FFT Algorithms
Cooley-Tukey
• Note
• Finally, the 4-point DFTs of the rows are computed:

• Output 2-D array is given by

• The process illustrated is precisely the first stage of the DIF

FFT algorithm

Original PowerPoint slides prepared by S. K. Mitra 11-61

© The McGraw-Hill Companies, Inc., 2007
 Cooley Tukey FFT Algorithms
Cooley-Tukey
• By choosing N1 = 4 and N2 = 2 , we get the first stage of the
DIT FFT algorithm
l ith
• Alternate index mappings are given by

Original PowerPoint slides prepared by S. K. Mitra 11-62

© The McGraw-Hill Companies, Inc., 2007
 Prime Factor Algorithms
• Twiddle factors can be eliminated by defining the index
mappings
i as

• To eliminate the twiddle factors we need to express

• Now
N

• It follows from above that if

‫ۃ‬AC‫ۄ‬N = N2, ‫ۃ‬BD‫ۄ‬N = N1, ‫ۃ‬AD‫ۄ‬N = ‫ۃ‬BC‫ۄ‬N = 0
then
Original PowerPoint slides prepared by S. K. Mitra 11-63
© The McGraw-Hill Companies, Inc., 2007
 Prime Factor Algorithms
• One set of coefficients that eliminates the twiddle factors is
given
i b
by
A = N2, B = N1, C = N2‫ۃ‬N2−1‫ۄ‬N1, D = N1‫ۃ‬N1−1‫ۄ‬N2
• Here ‫ۃ‬N1−1‫ۄ‬N2 denotes the multiplicative in erse of N1
m ltiplicati e inverse
reduced modulo N2
• Ö If ‫ۃ‬N1−11‫ۄ‬N2 = α then ‫ۃ‬N1α‫ۄ‬N2 = 1 or,
or in other words N1α =
N2β + 1 where β is any integer
example if N1 = 4 and N2 = 3 , then ‫ۃ‬3−1‫ۄ‬4 = 3 and
• For example,
‫ۃ‬3‫ڄ‬3‫ۄ‬4 = 1
• Likewise, if ‫ۃ‬N2−1‫ۄ‬N1 = γ, then N2γ = N1δ + 1 where δ is any
integer

Original PowerPoint slides prepared by S. K. Mitra 11-64

© The McGraw-Hill Companies, Inc., 2007
 Prime Factor Algorithms
• Now, ‫ۃ‬AC‫ۄ‬N = ‫ۃ‬N2‫(ڄ‬N2‫ۃ‬N2−1‫ۄ‬N1)‫ۄ‬N = ‫ۃ‬N2(N1δ+1)‫ۄ‬N =
‫ۃ‬N2N1δ+N N2)‫ۄ‬N = N2
• Similarly, ‫ۃ‬BD‫ۄ‬N = ‫ۃ‬N1‫(ڄ‬N1‫ۃ‬N1−1‫ۄ‬N2)‫ۄ‬N = ‫ۃ‬N1(N2δ+1)‫ۄ‬N =
‫ۃ‬N1N2δ+N1)‫ۄ‬N = N1
• Next,
‫ۃ‬AD‫ۄ‬N = ‫ۃ‬N2‫(ڄ‬N
(N1‫ۃ‬N1−11‫ۄ‬N2)‫ۄ‬N = ‫ۃ‬Nα‫ۄ‬
‫ۃ‬N ‫ۄ‬N = 0
• Likewise,
‫ۃ‬BC‫ۄ‬N = ‫ۃ‬N1‫(ڄ‬N
(N2‫ۃ‬N2−11‫ۄ‬N1)‫ۄ‬N = ‫ۃ‬Nγ‫ۄ‬
‫ۃ‬N ‫ۄ‬N = 0
• Hence,

Original PowerPoint slides prepared by S. K. Mitra 11-65

© The McGraw-Hill Companies, Inc., 2007
 Prime Factor Algorithms
• Thus,

where

and 0 ≤ k1 ≤ N1 −1 and 0 ≤ k2 ≤ N2 −1

Original PowerPoint slides prepared by S. K. Mitra 11-66

© The McGraw-Hill Companies, Inc., 2007
 Prime Factor Algorithms
• Example: Let N = 12. Choose N1 = 4 and N2 = 3
• Then, A = 3, B = 4, C = 3‫ۃ‬3−1‫ۄ‬4 = 9, and D = 4‫ۃ‬4−1‫ۄ‬3 = 4
• The index mappings are

• 2-D array representation of input is

Original PowerPoint slides prepared by S. K. Mitra 11-67

© The McGraw-Hill Companies, Inc., 2007
 Prime Factor Algorithms
• 4-point transforms of the columns lead to

• Final DFT array is

• 4-point DFTs require no multiplications, whereas the 3-point

DFTs require 4 complex multiplications
• Thus, the algorithm requires 16 complex multiplications 11-68
Original PowerPoint slides prepared by S. K. Mitra
© The McGraw-Hill Companies, Inc., 2007
 Chirp z-Transform
z Transform Algorithm
• Let x[n] be a length-N sequence with a Fourier transform
• We consider evaluation of M samples of that are equally
spaced in angle on the unit circle at frequencies
ωk = ωo + kΔω, 0 ≤ k ≤ M −1
where the starting frequency ωo and the frequency
i
increment t Δω
Δ can beb chosen
h arbitrarily
bit il

Original PowerPoint slides prepared by S. K. Mitra 11-69

© The McGraw-Hill Companies, Inc., 2007
 Chirp z-Transform
z Transform Algorithm
• The problem is thus to evaluate

or with
or, ith W defined as
W = e−jΔω
T evaluate
To l t

• Using
U i ththe id
identity
tit nkk = ½ [n
[ 2 + k2 − (k−n)
(k )2] we can write
it

• Letting
• We arrive at
Original PowerPoint slides prepared by S. K. Mitra 11-70
© The McGraw-Hill Companies, Inc., 2007
 Chirp z-Transform
z Transform Algorithm
• Interchanging k and n we get

• Th
Thus,
s X(e−jωn) corresponds to the con
convolution
ol tion of the
sequence g[n] with the sequence W−n2/2 followed by
multiplication by the sequence Wn2/2 as indicated below
• We arrive at

• The sequence can be thought of as a complex exponential

sequence with linearly increasing frequency
• Such signals, in radar systems, are called chirp signals 11-71
Original PowerPoint slides prepared by S. K. Mitra
© The McGraw-Hill Companies, Inc., 2007
 Chirp z-Transform
z Transform Algorithm
• For the evaluation of

the output of the system depicted earlier need to be

computed over a finite interval
• Since g[n] is a length-N sequence, only a finite portion of the
infinite length sequence W−n2/2 is used in obtaining the
convolution sum over the interval 0 ≤ n ≤ M −1 1

Original PowerPoint slides prepared by S. K. Mitra 11-72

© The McGraw-Hill Companies, Inc., 2007
 Chirp z-Transform
z Transform Algorithm
• Typical signals

Original PowerPoint slides prepared by S. K. Mitra 11-73

© The McGraw-Hill Companies, Inc., 2007
 Chirp z-Transform
z Transform Algorithm
• The portion of the sequence W−n2/2 used in obtaining the
convolution
l ti sum iis ffrom th
the iinterval
t l −N
N +1
1 ≤ n ≤ M −1
1

• Let

as shown below

• It can be seen that

Original PowerPoint slides prepared by S. K. Mitra 11-74

© The McGraw-Hill Companies, Inc., 2007
 Chirp z-Transform
z Transform Algorithm
• Hence, the computation of the frequency samples X(ejωn)
can be
b carried
i d outt using
i an FIR filt
filter as iindicated
di t d bbelow
l

where y[n] = X(ejωn), 0 ≤ n ≤ M −1

• Advantages
– N = M is not required as in FFT algorithms
– Neither N nor M do not have to be composite numbers
– Parameters ωo and Δω are arbitrary
– Convolution with h[n] can be implemented using FFT
techniques
Original PowerPoint slides prepared by S. K. Mitra 11-75
© The McGraw-Hill Companies, Inc., 2007
 Chirp z-Transform
z Transform Algorithm

Original PowerPoint slides prepared by S. K. Mitra 11-76

© The McGraw-Hill Companies, Inc., 2007