Fast Fourier Transform (FFT) (Theory and Implementation)

Chapter 19
Fast Fourier Transform (FFT)

(Theory and Implementation)
Chapter 19, Slide 1
Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004
Learning Objectives
Chapter 19, Slide 2
DFT algorithm.
Conversion of DFT to FFT algorithm.
Implementation of the FFT algorithm.
DFT Algorithm
The Fourier transform of an analogue

signal x(t) is given by:
X
x t e jt dt
The Discrete Fourier Transform (DFT) of

a discrete-time signal x(nT) is given by:
N 1
X k x n e
2
nk
N
n 0
Chapter 19, Slide 3
Where:
k 0,1, N 1
x nT x n
DFT Algorithm
N 1
X k x nWNnk
n0
x[n] = input
X[k] = frequency bins
W
= twiddle factors
X(0)
= x[0]WN0 + x[1]WN0*1 ++ x[N-1]WN0*(N-1)
X(1)
= x[0]WN0 + x[1]WN1*1 ++ x[N-1]WN1*(N-1)

:
X(k)
= x[0]WN0 + x[1]WNk*1 ++ x[N-1]WNk*(N-1)

:
X(N-1) = x[0]WN0 + x[1]WN (N-1)*1 ++ x[N-1]WN (N-1)(N-1)
Note: For N samples of x we have N frequencies

representing the signal.
Chapter 19, Slide 4
Performance of the DFT Algorithm
The DFT requires N2 (NxN) complex

multiplications:
The DFT also requires (N-1)*N complex

additions:
Chapter 19, Slide 5
Each X(k) requires N complex

multiplications.
Therefore to evaluate all the values of the
DFT ( X(0) to X(N-1) ) N2 multiplications are
required.
Each X(k) requires N-1 additions.

Therefore to evaluate all the values of the
DFT (N-1)*N additions are required.
120
120
Number of Additions
Number of Multiplications
Performance of the DFT Algorithm
100
80
60
40
20
100
80
60
40
20
0
0
0
Number of Samples
Chapter 19, Slide 6
10
10
Number of Samples
Can the number of computations required

be reduced?
DFT FFT
Chapter 19, Slide 7
A large amount of work has been devoted

to reducing the computation time of a
DFT.
This has led to efficient algorithms which
are known as the Fast Fourier Transform
(FFT) algorithms.
DFT FFT
N 1
X k x nWNnk ; 0 k N 1
n 0
[1]
x[n] = x[0], x[1], , x[N-1]
Lets divide the sequence x[n] into even

and odd sequences:
Chapter 19, Slide 8
x[2n]
= x[0], x[2], , x[N-2]
x[2n+1] = x[1], x[3], , x[N-1]
DFT FFT
Equation 1 can be rewritten as:

N
1
2
N
1
2
n 0
n0
X k x 2nWN2 nk x 2n 1WN 2 n 1 k
[2]
Since:
W N2 nk e
2
2 nk
N
2
nk
N 2
WN 2 n 1 k WNk W Nnk
W Nnk
Then:
X k
N
1
2
n 0
x 2nW Nnk W Nk
2
Y k WNk Z k
Chapter 19, Slide 9
N
1
2
n 0
x 2n 1W Nnk
2
DFT FFT
The result is that an N-point DFT can be

divided into two N/2 point DFTs:
N 1
X k x nWNnk ; 0 k N 1
N-point DFT
n 0
Where Y(k) and Z(k) are the two N/2

point DFTs operating on even and odd
samples respectively:
X k
N
1
2
n 0
x1 nW Nnk W Nk
2
Y k W Nk Z k
Chapter 19, Slide 10
N
1
2
n 0
x2 nW Nnk
2
Two N/2point DFTs
DFT FFT
Periodicity and symmetry of W can be

exploited to simplify the DFT further:
N
1
2
X k
x1 n W Nnk W Nk
n 0
N
1
2
x2 nW Nnk
n 0
[3]
X k
2
Or:
And:
WN
N
2
N
1
2
N
n k
x1 n W N 2
n 0
2
WN
2
N
2
2
k
N
2 N
N 2
2
k
N 2
N
1
2
N
k
WN 2
x2
n 0
2
k
N
2 N
N 2 2
e
2
k
N 2
n k
n W N 2
2
2
k
N
W Nk
2
W Nk
: Symmetry
: Periodicity
DFT FFT
Symmetry and periodicity:

W86
W85
W87
W84
W80 = W88
W83
W81 = W89
WNk+N/2 = -WNk
WN/2k+N/2 = WN/2k
W8k+4 = -W8k
W8k+8 = W8k
W82
DFT FFT
Finally by exploiting the symmetry and

periodicity, Equation 3 can be written as:
N
X k
2
N
1
2
n 0
x1 nW Nnk WNk
2
Y k WNk Z k
N
1
2
n 0
x2 nW Nnk
2
[4]
DFT FFT
N
1
2
N
k
Y k WN Z k ; k 0,
1
2
X k Y k WNk Z k ; k 0,
N
X k
2
Y(k) and WNk Z(k) only need to be

calculated once and used for both
equations.
Note: the calculation is reduced from 0 to
N-1 to 0 to (N/2 - 1).
DFT FFT
N
1
2
N
k
Y k WN Z k ; k 0,
1
2
X k Y k WNk Z k ; k 0,
N
X k
2
Y(k) and Z(k) can also be divided into N/4

point DFTs using the same process shown
above:
Y k U k W Nk V k
2
N
Y k U k W Nk V k
4
Z k P k W Nk Q k
2
Z k P k W Nk Q k
4
The process continues until we reach 2

point DFTs.
DFT FFT
x[0]
x[2]
x[4]
N/2 point
DFT
x[N-1]
X[0] = y[0]+W0z[0]
X[1] = y[1]+W1z[1]
y[N-2]
x[N-2]
x[1]
x[3]
x[5]
y[0]
y[1]
y[2]
N/2 point
DFT
z[0]
z[1]
z[2]
W0
W1
-1
-1
X[N/2] = y[0]-W0z[0]
X[N/2+1] = y[1]-W1z[1]
z[N/2-1]
Illustration of the first decimation in time

FFT.
X [k ] G[k] WNk H[ K ]
X[k
N
] G[k] WNk H[ K ]
2
N
k 0 to
1
2
G[k ] G1[k] W Nk H1[ K ]

2
N
G[k ] G1[k] W Nk H1[ K ]
4
2
N
k 0 to ( 1)
4
BUTTERFLY STRUCTURE
Decimation-In-Time FFT Algorithm
Final flow graph for 8-point decimation in time
FFT Implementation
To efficiently implement the FFT

algorithm a few observations are made:
Each stage has the same number of

butterflies (number of butterflies = N/2, N is
number of points).
The number of DFT groups per stage is equal
to (N/2stage).
The difference between the upper and lower
leg is equal to 2stage-1.
The number of butterflies in the group is
equal to 2stage-1.
FFT Implementation
Example: 8 point FFT
W0
W0
W0
W0
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1
Decimation in time FFT:
Number of stages = log2N

Number of blocks/stage = N/2stage
Number of butterflies/block
= 2stage-1
Dr. Naim Dahnoun,
Bristol University, (c) Texas Instruments 2004
FFT Implementation
(1) Number of stages:
W0
W0
W0
W0
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
W0
W0
W0
W0
FFT Implementation
Nstages = 1
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
W0
W0
W0
W0
FFT Implementation
Stage 2

Nstages = 2
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
W0
W0
W0
W0
FFT Implementation
Stage 2
Stage 3
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
W0
W0
W0
W0
FFT Implementation
Stage 2
Stage 3
-1
-1
-1
-1
W0
-1
W2
-1
W0

Nstages = 3
(2) Blocks/stage:
Stage 1:
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
FFT Implementation
Stage 2
Stage 3
Block 1
W0
W0
W0
W0
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:
Stage 1: Nblocks = 1

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
FFT Implementation
Stage 2
Stage 3
Block 1
W0
-1
Block 2
W0
W0
W0
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
FFT Implementation
Stage 2
Stage 3
Block 1
W0
-1
Block 2
W0
-1
W0
-1
W2
-1
Block 3
W0
W0
-1
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
FFT Implementation
Stage 2
Stage 3
Block 1
W0
-1
Block 2
W0
-1
W0
-1
W2
-1
Block 3
W0
-1
Block 4
W0
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
Block 1
W0
W0
W0
W0
FFT Implementation
Stage 2
Stage 3
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
Block 1
W0
W0
Block 2
W0
W0
FFT Implementation
Stage 2
Stage 3
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
W0
Block 1
W0
W0
W0
FFT Implementation
Stage 2
Stage 3
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
W0
W0
W0
W0
FFT Implementation
Stage 2
Stage 3
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:
(3) Bflies/block:
Stage 1:

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
W0
W0
W0
W0
FFT Implementation
Stage 2
Stage 3
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:
(3) Bflies/block:
Stage 1: Nbtf = 1

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
W0
W0
W0
W0
FFT Implementation
Stage 2
Stage 3
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:
(3) Bflies/block:
Stage 1: Nbtf = 1
Stage 2: Nbtf = 1

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
W0
W0
W0
W0
FFT Implementation
Stage 2
Stage 3
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:
(3) Bflies/block:
Stage 1: Nbtf = 1
Stage 2: Nbtf = 2

= 2stage-1
Dr. Naim Dahnoun,
Stage 1
W0
W0
W0
W0
FFT Implementation
Stage 2
Stage 3
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:
(3) Bflies/block:
Stage 1: Nbtf = 1
Stage 2: Nbtf = 2

Stage 3: Nbtf = 1
= 2stage-1
Dr. Naim Dahnoun,
Stage 1
W0
W0
W0
W0
FFT Implementation
Stage 2
Stage 3
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:
(3) Bflies/block:
Stage 1: Nbtf = 1
Stage 2: Nbtf = 2

Stage 3: Nbtf = 2
= 2stage-1
Dr. Naim Dahnoun,
Stage 1
W0
W0
W0
W0
FFT Implementation
Stage 2
Stage 3
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:
(3) Bflies/block:
Stage 1: Nbtf = 1
Stage 2: Nbtf = 2

Stage 3: Nbtf = 3
= 2stage-1
Dr. Naim Dahnoun,
Stage 1
W0
W0
W0
W0
FFT Implementation
Stage 2
Stage 3
-1
-1
-1
-1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1

Nstages = 3
(2) Blocks/stage:
(3) Bflies/block:
Stage 1: Nbtf = 1
Stage 2: Nbtf = 2

Stage 3: Nbtf = 4
= 2stage-1
Dr. Naim Dahnoun,
FFT Implementation
Stage 1
W0
-1
W0
-1
W0
Stage 3
-1
W0
Start Index
Input Index
Twiddle Factor Index
Stage 2
-1
0
1
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1
0
2
0
4
N/2 = 4
FFT Implementation
Stage 1
W0
-1
W0
-1
W0
Stage 3
-1
W0
Start Index
Input Index
Stage 2
-1
0
1
N/2 = 4
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1
0
2
0
4
4 /2 = 2
FFT Implementation
Stage 1
W0
-1
W0
-1
W0
Stage 3
-1
W0
Start Index
Input Index
Stage 2
-1
0
1
N/2 = 4
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1
0
2
4 /2 = 2
0
4
2 /2 = 1
FFT Implementation
Stage 1
W0
Stage 3
-1
W0
-1
W0
-1
W0
Start Index
Input Index
Indicies Used
Stage 2
-1
0
1
N/2 = 4
W0
W0
-1
W2
-1
W0
W0
-1
W1
-1
W0
-1
W2
-1
W2
-1
W3
-1
0
2
0
4
4 /2 = 2
W0
2 /2 = 1
W2
W1
W0
W2
Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments

2004
3
Decimation-in-frequency FFT
Regroup the signal samples
butterfly operation
DFT obtained with proper order of frequency index
Assuming that N is a even number.
Divide the data sample into two groups
x (0), x (1), x ( 2)... x ( 2 1)

N
N
x
(
),
x
(
1
),...
x
(
N
1
)
.
2
2
7
X ( k ) x ( n )W
n 0
nk
8
x ( n )W
n 0
nk
8
x ( n )W
n 4
nk
8
We want to do something so that the above

are two 4-point DFTs.
For even k, (k=2m)
3
X ( 2m ) x ( n )W
2 nm
8
n 0
x ( n )W
n 0
nm
4
x ( n )W82 nm
n 4
x ( n 4)W
n 0
( n4)m
4
x ( n )W
n 0
nm
4
x ( n 4)W4nm
n 0
x ( n ) x ( n 4)W4nm
n 0
The DFT with even frequency indices can be

obtained as the 4-DFT of [x(n) +x(n+4)] (n=0,1,2,3)
For odd k, (k=2m+1)
X ( 2m 1)
x ( n )W
n ( 2 m 1)
8
n 0
W
n 0
W
n 0
n
8
n
8
W
n 0
n
8
nm
4
x ( n )W
nm
4
x ( n )W
W
n 0
( n4 )
8
W
n 0
n
8
x ( n )W
n ( 2 m 1)
8
n 4
x ( n 4)W4( n 4 ) m
x ( n 4)W4nm
x ( n ) x ( n 4)W
nm
4
W x ( n ) x ( n 4)W
3
n 0
n
8
nm
4
The DFT with odd frequency

indices can be obtained
n
as the 4-DFT of W8 x ( n ) x ( n 4)
(n=0,1,2,3)
x(0)
x(1)
x(2)
x(3)
x(4)
x(5)
x(6)
x(7)
X(0)
4-point
DFT
-
W80
W81
4-point
W82
DFT
W83
X(2)
X(4)
X(6)
X(1)
X(3)
X(5)
X(7)
x(0)
2-point
x(1)
x(2)
x(3)
x(4)
x(5)
x(6)
x(7)
W40
W41
W80
1
8
W82
W83
W40
W41
X(0)
DFT
X(4)
2-point
X(2)
DFT
X(6)
2-point
X(1)
DFT
X(5)
2-point
X(3)
DFT
X(7)
x(0)
X(0)
x(1)
x(2)
x(3)
x(4)
x(5)
x(6)
x(7)
--
W80
W81
W82
W83
W40
W41
W20
X(4)
X(2)
X(6)
X(1)
X(3)
W40
W41
X(5)
X(7)
End

Fast Fourier Transform (FFT) (Theory and Implementation)

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Fast Fourier Transform (FFT) (Theory and Implementation)

Încărcat de

Drepturi de autor:

Formate disponibile

Chapter 19

Fast Fourier Transform (FFT)

Chapter 19, Slide 1

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

Chapter 19, Slide 2

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

The Fourier transform of an analogue

The Discrete Fourier Transform (DFT) of

Chapter 19, Slide 3

= x[0]WN0 + x[1]WN0*1 ++ x[N-1]WN0*(N-1)

= x[0]WN0 + x[1]WN1*1 ++ x[N-1]WN1*(N-1)

= x[0]WN0 + x[1]WNk*1 ++ x[N-1]WNk*(N-1)

X(N-1) = x[0]WN0 + x[1]WN (N-1)*1 ++ x[N-1]WN (N-1)(N-1)

Note: For N samples of x we have N frequencies

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

Performance of the DFT Algorithm

The DFT requires N2 (NxN) complex

The DFT also requires (N-1)*N complex

Chapter 19, Slide 5

Each X(k) requires N complex

Each X(k) requires N-1 additions.

Performance of the DFT Algorithm

Chapter 19, Slide 6

Can the number of computations required

Chapter 19, Slide 7

A large amount of work has been devoted

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

x[n] = x[0], x[1], , x[N-1]

Lets divide the sequence x[n] into even

Chapter 19, Slide 8

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

Equation 1 can be rewritten as:

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

The result is that an N-point DFT can be

Where Y(k) and Z(k) are the two N/2

Two N/2point DFTs

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

Periodicity and symmetry of W can be

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

Symmetry and periodicity:

Chapter 19, Slide 12

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

Finally by exploiting the symmetry and

Chapter 19, Slide 13

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

Chapter 19, Slide 14

Y(k) and WNk Z(k) only need to be

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

Y(k) and Z(k) can also be divided into N/4

Chapter 19, Slide 15

The process continues until we reach 2

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

Chapter 19, Slide 16

Illustration of the first decimation in time

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

Chapter 19, Slide 18

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

G[k ] G1[k] W Nk H1[ K ]

Chapter 19, Slide 19

Chapter 19, Slide 20

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

Chapter 19, Slide 21

Dr. Naim Dahnoun, Bristol University, (c) Texas Instruments 2004

= x[0]WN0 + x[1]WN01 ++ x[N-1]WN0(N-1)

= x[0]WN0 + x[1]WN11 ++ x[N-1]WN1(N-1)

= x[0]WN0 + x[1]WNk1 ++ x[N-1]WNk(N-1)