Biomedical

Chapter Signal processing

9 Computation of the
Discrete
Fourier Transform
Zhongguo Liu
Biomedical Engineering
School of Control Science and Engineering, Shandong
University

Zhongguo Liu_Biomedical Engineering_Shandong

25/4/27 1
1 Univ.
 Chapter 9 Computation of the
Discrete Fourier Transform
9.0 Introduction
.1 Efficient Computation of Discrete Fourier Transfor
.2 The Goertzel Algorithm
.3 decimation-in-time FFT Algorithms
.4 decimation-in-frequency FFT Algorithms
.5 practical considerations （ software realization)

2
 9.0 Introduction
Implement a convolution of two
sequences by the following procedure:
1. Compute the N-point DFTX 1 k  X 2 k 
and
x1 n
of the two sequence x2 nand
2. ComputeX 3 k   X 1 k X 2 k  for0 k  N  1
3. Computex3 n  x1 n N x2 n as the
X 3 k  of
inverse DFT
Why not convolve the two sequences directly?

There are efficient algorithms called Fast

Fourier Transform (FFT) that can be orders
of magnitude more efficient than others.
3
 9.1 Efficient Computation of
Discrete Fourier Transform
The DFT pair was given as
N1
 j 2 / N kn 1 N1
j 2 / N kn
X  k   x[n]e x[n]   X k e
n 0
N k 0

Baseline for computational complexity:

Each DFT coefficient requires
N complex multiplications;
N-1 complex additions
All N DFT coefficients require
N2 complex multiplications;
N(N-1) complex additions
4 4
 9.1 Efficient Computation of
Discrete Fourier Transform
N1
 j 2 / N kn
X  k   x[n]e
n 0

Complexity in terms of real operations

4N2 real multiplications
2N(N-1) real additions (approximate 2N2)

5 5
 9.1 Efficient Computation of
Discrete Fourier Transform
Most fast methods are based on Periodicity properties
Periodicity in n and k; Conjugate symmetry
 j 2 / N k  N  n   j 2 / N kN  j 2 / N k  n  j 2 / N kn
e e e e
 j 2 / N kn  j 2 / N k n  N  j 2 / N k  N n
e e e
Re   ]

6 6
 9.2 The Goertzel Algorithm
Makes use of the periodicity j 2 / N Nk
e e j 2 k 1
Multiply DFT equation with this factor
j 2 / N kN
N1
 j 2 / N rk N1
j 2 / N k  N  r 
X  k  e  x[r ]e  x[r ]e
r 0 r 0

Define

j 2 / N k n  r 
yk  n    x[r ]e u n  r
r  
using x[n]=0 for n<0 and n>N-1
X  k   yk  n  n N
X[k] can be viewed as the output of a filter to the input x[n]
Impulse response of filter: j 2 / N kn
h[n] e u  n
X[k] is the output of the filter at time n=N
7 7
 9.2 The Goertzel Algorithm
Goertzel j 2 / N kn
Filter: h[n] e u[n] W  knu[n]
N

1
H k z  
1  WN k z  1

yk [n]  yk [n  1]WN k  x[n], n 0,1,..., N , yk [ 1] 0

X  k   yk  n  n N , k 0,1,..., N
N1
X  k   x[n]WNkn
n 0
Computational complexity
4N real multiplications; 4N real additions
Slightly less efficient than the direct method
But it avoids computation and storage ofW kn
8
N 8
 Second Order Goertzel Filter
Goertzel Filter
1
H k z   2
j k 1
1 e N z
Multiply both numerator and denominator
2 2
 j k  j k
1 e N
z 1 1 e N
z 1
H k z   
 j
2
k   2
 j k 1  2 k  1  2
 1  e N
z  1
 1  e N
z  1  2 cos z z
N
  
2 k
y[n]  y[n  2]  2 cos y[n  1]  x[n], n 0,1,..., N
N
yk [ N ]  y[ N ]  WNk y[ N  1]  X  k  , k 0,1, ..., N
9 9
 Second Order Goertzel Filter
2 k
y[n]  y[n  2]  2 cos y[n  1]  x[n], n 0,1,..., N
N
yk [ N ]  y[ N ]  WNk y[ N  1]  X  k  , k 0,1, ..., N
Complexity for one DFT coefficient ( x(n) is complex
sequence).
Poles: 2N real multiplications and 4N real additions
Zeros: Need to be implement only once:
4 real multiplications and 4 real additions
Complexity for all DFT coefficients
Each pole is used for two DFT coefficients
Approximately N2 real multiplications and 2N2 real
additions
1 10
 Second Order Goertzel Filter

2 k
y[n]  y[n  2]  2 cos y[n  1]  x[n], n 0,1,..., N
N

yk [ N ]  y[ N ]  WNk y[ N  1]  X  k  , k 0,1, ..., N

If do not need to evaluate all N DFT coefficients

Goertzel Algorithm is more efficient than FFT if
less than M DFT coefficients are needed,M < log2N

11 11
 9.3 decimation-in-time FFT
Algorithms
Makes use of both periodicity and symmetry
Consider special case of N an integer power of 2
Separate x[n] into two sequence of length N/2
Even indexed samples in the first sequence
Odd indexed samples in the other sequence

N1
 j 2 / N kn
X  k   x[n]e
n 0

 j 2 / N kn  j 2 / N kn

  x[n]e   x[n]e
n even n odd
1 12
 9.3 decimation-in-time FFT
Algorithms
 j 2 / N kn  j 2 / N kn
X  k    x[n]e   x[n]e
n even n odd
Substitute variables n=2r for n even and n=2r+1 for odd
N /2  1 N /2  1
X  k    x[2r ]W 2 rk
N   x[2r  1]WN 2 r 1k
r 0 r 0
N /2  1 N /2  1
  x[2r ]WNrk/2  WNk  x[2r  1]WNrk/2
r 0 r 0

G  k   W H  k 
k  j 2 2  j 2
N W 2
N e N e N /2 WN /2
G[k] and H[k] are the N/2-point DFT’s of each subsequence
1 13
 9.3 decimation-in-time FFT
Algorithms
N /2  1 N /2  1
X  k    x[2r ]W rk
N /2
k
W
N  x[2r  1]W rk
N /2
r 0 r 0

G  k   W H  k 
k
 j 2 2 rk  j 2 rk
N
e N e N /2 WNrk/2
N1
k 0,1,..., k 0,1,..., N
2
 N  N
G  k   G  k  H  k   H  k 
 2  2
G[k] and H[k] are the N/2-point DFT’s of each subsequence

1 14
 8-point DFT using decimation-in-time

1 Figure 9.3
 computational complexity
Two N/2-point DFTs
2(N/2)2 complex multiplications
2(N/2)2 complex additions
Combining the DFT outputs
N complex multiplications
N complex additions
Total complexity
N2/2+N complex multiplications
N2/2+N complex additions
More efficient than direct DFT

1 16
 9.3 decimation-in-time FFT
Algorithms

Repeat same process ,

Divide N/2-point DFTs into
Two N/4-point DFTs
Combine outputs

N=8

1 17
 9.3 decimation-in-time FFT
Algorithms
After two steps of decimation in time

Repeat until we’re left with two-point DFT’s

1 18
 9.3 decimation-in-time FFT
Algorithms
flow graph for 8-point decimation in time

Complexity:
Nlog2N complex multiplications and additions
1 19
 Butterfly Computation
Flow graph constitutes of butterflies

We can implement each butterfly with one multiplication

Final complexity for decimation-in-time FFT

(N/2)log2N complex multiplications and additions
2 20
 9.3 decimation-in-time FFT
Algorithms
Final flow graph for 8-point decimation in time

Complexity:
(Nlog2N)/2 complex multiplications and Nlog2N additions
2 21
 9.3.1 In-Place Computation 同址运算
Decimation-in-time flow graphs require two sets of registers
Input and output for each stage

X 0  0  x  0 x  0 X 2  0 X  0
X 0  1  x  4 x  4 X 2  1 X 1
X 0  2  x  2 x  2 X 2  2 X  2
X 0  3  x  6 x  6 X 2  3 X  3
X 0  4  x 1 x 1 X 2  4 X  4
X 0  5  x  5 x  5 X 2  5 X  5
X 0  6  x  3 x  3 X 2  6 X  6
2X0  7 x  7 x  7 X 2  7 X  7 22
 9.3.1 In-Place Computation 同址运算
Note the arrangement of the input indices
Bit reversed indexing （码位倒置）

X 0  0  x  0  X 0  000  x  000 x  0 X  0
X 0  1  x  4  X 0  001  x 100 x  4 X 1
X 0  2  x  2  X 0  010  x  010 x  2 X  2
X 0  3  x  6  X 0  011  x 110 x  6 X  3
X 0  4  x 1  X 0 100  x  001 x 1 X  4
X 0  5  x  5  X 0 101 x 101 x  5 X  5
X 0  6  x  3  X 0 110  x  011 x  3 X  6
X 0  7   x  7   X 0 111  x 111 x  7 X  7

2 23
 cause of bit-reversed order

binary coding for position ：

000
001

010
011

100
101

110
111

must padding 0 to Figure 9.13

N 2 M
2
 9.3.2 Alternative forms
Note the arrangement of the input indices
Bit reversed indexing （码位倒置）

X 0  0  x  0  X 0  000  x  000 x  0 X  0
X 0  1  x  4  X 0  001  x 100 x  4 X 1
X 0  2  x  2  X 0  010  x  010 x  2 X  2
X 0  3  x  6  X 0  011  x 110 x  6 X  3
X 0  4  x 1  X 0 100  x  001 x 1 X  4
X 0  5  x  5  X 0 101 x 101 x  5 X  5
X 0  6  x  3  X 0 110  x  011 x  3 X  6
X 0  7   x  7   X 0 111  x 111 x  7 X  7

2 25
9.3.2 Alternative forms

strongpoint ： in-place computations

shortcoming ： non-sequential access of data
Figure 9.14
2
 Figure 9.15

shortcoming ： not in-place computation

non-sequential access of data
2
 Figure 9.16

shortcoming ： not in-place computation

strongpoint: sequential access of data
2
 9.3 decimation-in-time FFT
Algorithms
 j 2 / N kn  j 2 / N kn
X  k    x[n]e   x[n]e
n even n odd
Substitute variables n=2r for n even and n=2r+1 for odd
N /2  1 N /2  1
X  k    x[2r ]W 2 rk
N   x[2r  1]WN 2 r 1k
r 0 r 0
Review
N /2  1 N /2  1
  x[2r ]WNrk/2  WNk  x[2r  1]WNrk/2
r 0 r 0

G  k   W H  k 
k  j 2 2  j 2
N W 2
N e N e N /2 WN /2
G[k] and H[k] are the N/2-point DFT’s of each subsequence
2 29
 9.3.1 In-Place Computation 同址运算
Bit reversed indexing （码位倒置）

X 0  000 x  000 x  0 X  0
X 0  001 x 100 x  4 X 1
X 0  010 x  010 x  2 X  2
X 0  011 x 110 x  6 X  3
X 0 100 x  001 x 1 X  4
X 0 101 x 101 x  5 X  5
X 0 110 x  011 x  3 X  6
X 0 111 x 111 x  7  X  7

3 30
9.3.2 Alternative forms

strongpoint ： in-place computations

shortcoming ： non-sequential access of data
Figure 9.14
3
 9.4 Decimation-In-Frequency FFT
Algorithm
N1
The DFT equation X  k   x[n]W nk
N
n 0
Split the DFT equation into even and odd frequency indexes

N1 N / 2 1 N1
X  2r   x[n]WNn 2 r   x[n]WNn 2 r   x[n]WNn 2 r
n 0 n 0 n N / 2
N /2  1 N /2  1
Substitute   x[n]W n  N /2 2 r
variables n 0
n2r
N  
n 0
x[n  N / 2]WN
N /2  1
   x[n]  x[n  N / 2]W nr
N /2
n 0
N /2  1
  g (n)WNrn/2
n 0
3 32
 9.4 Decimation-In-Frequency FFT
Algorithm N1
The DFT equation X  k   x[n]WNnk
n 0
N1 N /2  1 N1
X  2r  1  x[n]W n (2 r 1)
N   x[n]W n (2 r 1)
N   x[n]W n (2 r 1)
N
n 0 n 0 n N /2
N /2  1 N /2  1
n  N /2 (2 r 1)
  x[n]W n (2 r 1)
N   x[n  N / 2]WN
n 0 n 0
N /2  1
   x[n]  x[n  N / 2]W n (2 r 1)
N
n 0
N / 2 1 N /2  1
   x[n]  x[n  N / 2]W W n
N
rn
N /2
  h(n)WNnWNrn/2
n 0 n 0
N
n 2 r 1 2 rn n rn n (2 r 1)
W N W W W W
N N N /2 N W N
2
WNNrWNN /2  1
3 33
 decimation-in-frequency decomposition of an N-point
DFT to N/2-point DFT

N /2  1 N /2  1
X  2r     x[n]  x[n  N / 2]WNnr/2  g (n)WNrn/2
n 0N /2  1 n 0 N /2  1
X  2r  1    x[n]  x[n  N / 2]WNnW
rn
N /2
  h(n)WNnWNrn/2
3 n 0 n 0 34
 decimation-in-frequency decomposition of an 8-point
DFT to four 2-point DFT

N / 4 1 N /4  1
X  2* 2 s    [ g (n)  g (n  N / 4)]WNsn/4   p (n)WNsn/4
n 0 n 0
N / 4 1 N /4  1
X  2*(2 s  1)    [ g (n)  g (n  N / 4)]W W
2n
N
sn
N /4   q (n)WN2 nWNsn/4
3 n 0 n 0 35
 2-point DFT

X v ( p)  X v 1 ( p)  X v 1 (q)

X v (q )  X v  1 ( p )  X v  1 (q )  W80 when N 8

3 36
 N /2  1 N /2  1
X  2r     x[n]  x[n  N / 2] WNnr/2   g (n)WNrn/2
n 0 n 0

N /4  1 N /2  1
X  2* 2 s    g (n)WN2/2sn   g (n)WN2/2sn
n 0 n N /4
N /4  1 N /4  1
  g (n)WN2/2sn   g (n  N / 4)WN2/2s ( n  N /4)
n 0 n 0
N /4  1 N /4  1
  g (n)W sn
N /4   g (n  N / 4)W sn
N /4
n 0 n 0
N /4  1 N /4  1
  [ g (n)  g (n  N / 4)]W sn
N /4
  p (n)WNsn/4
n 0 n 0

3
 N /2  1 N /2  1
X  2r     x[n]  x[n  N / 2] nr
WN /2   g (n)WNrn/2
n 0 n 0
N /2  1
X  2*(2 s  1)    g (n)WN(2/2s 1) n
n 0
N /4  1 N /2  1
  g (n)WN(2/2s 1) n   g (n)WN(2/2s 1) n
n 0 n N / 4
N /4  1 N /4  1
  g (n)WNsn/4WNn/2   g (n  N / 4)WN(2/2s 1)( n  N /4)
n 0 n 0
N /4  1 N /4  1
  g (n)WNsn/4WN2 n   g (n  N / 4)WNsn/4WN2 nWN(2/ 2s 1) N /4
n 0 n 0
N /4  1 N /4  1

  [ g (n)  g (n  N / 4)]W W 2n sn   q (n)WN2 nWNsn/4

N N /4
n 0 n 0

3
WN(2/2s 1) N /4 WNsN/2/2WNN/2/4  1
 N /4  1
X  2* 2 s    p (n)WNsn/4
n 0

N /4  1
X  2* 2* 2t    p(n)W 2 tn
N /4
n 0
N /8  1 N /4  1
  p (n)W 2 tn
N /4   p (n)W 2 tn
N /4
n 0 n N / 8

N /8  1 N /8  1
  p (n)W 2 tn
N /4   p (n  N / 8)W 2 t ( n  N /8)
N /4
n 0 n 0

N /8  1
  [ p (n)  p (n  N / 8)]WNtn/8
n 0

 p (n)  p (n  1) when N 8
3
 N /4  1
X  2* 2 s    p (n)WNsn/4
n 0
N /4  1
X  2* 2*(2t  1)    p( n)WN(2/4t 1) n
n 0
N /8  1 N /4  1
  p (n)WN(2/4t 1) n   p (n)WN(2/ 4t 1) n
n 0 n N / 8
N /8  1 N /8  1
  p (n)WN(2/4t 1) n   p (n  N / 8)WN(2/4t 1)( n  N /8)
n 0 n 0
N /8  1 N /8  1
  p (n)WN2/4
tn
WNn/4   p (n  N / 8)WN2/4
tn
WNn/4WN(2/4t 1) N /8
n 0 n 0
N /8  1
  [ p (n)  p (n  N / 8)]WNtn/8WN4 n
n 0 WN(2/4t 1) N /8 WNtN/4/4WNN/4/8  1
[ p (n)  p (n  1)]W80 when N 8
4
 Final flow graph for 8-point DFT
decimation in frequency

4 41
 9.4.1 In-Place Computation 同址运算

DIF
FFT

DIT FFT

4 42
 9.4.1 In-Place Computation 同址运算

DIF
FFT

DIT FFT

4 43
 9.4.2 Alternative forms
decimation-in-frequecy Butterfly Computation

decimation-in-time Butterfly Computation

4 44
 The DIF FFT is the transpose of the DIT FFT

DIF
FFT

DIT FFT

4 45
 9.4.2 Alternative forms

DIF
FFT

DIT FFT

4
 9.4.2 Alternative forms

DIF
FFT

DIT FFT

4
 Figure 9.24 erratum
x  0
x  4
x  2
x  6
x 1
x  5
x  3
x  7
4
 9.4.2 Alternative forms

DIF
FFT

DIT FFT

4
 Chapter 9 HW
9.1, 9.2, 9.3,

Zhongguo Liu_Biomedical Engineering_Shandon

50
5 返 回25/4/27 上一页 下一页 g Univ.