18-491/691 Lecture #16

FAST FOURIER TRANSFORM

ALTERNATE IMPLEMENTATIONS

Richard M. Stern

Department of Electrical and Computer Engineering

Carnegie Mellon University
Pittsburgh, Pennsylvania 15213
Phone: +1 (412) 268-2535
FAX: +1 (412) 268-3890
[email protected]
http://www.ece.cmu.edu/~rms

March 24, 2021

 Introduction

n In our lecture last Monday we described and discussed the

basic decimation-in-time Cooley-Tuckey fast Fourier transform
algorithm for DFT sizes that are integer powers of 2 (radix 2)
n Today we will discuss some variations and extensions of the
basic FFT algorithm:
– Computation of the inverse FFT
– One further trivial efficiency
– Alternate forms of the FFT structure
– The decimation-in-frequency FFT algorithm
– FFT structures for DFT sizes that are not an integer power of 2

Carnegie
Mellon Slide 2 ECE Department
 Alternate FFT structures

n We developed the basic decimation-in-time (DIT) FFT structure

in the last lecture, but other forms are possible simply by
rearranging the branches of the signal flowgraph

n Some issues to consider:

– Natural or bit-reversed input and output?
– In-place computation?

Carnegie
Mellon Slide 3 ECE Department
 Alternate DIT FFT structures (continued)

n DIT structure with input bit-reversed, output natural (OSYP

9.11):

Carnegie
Mellon Slide 4 ECE Department
 Alternate DIT FFT structures (continued)

n The original DIT n Rearranged structure

structure (OSYP 9.11): (OSYP 9.15):

Carnegie
Mellon Slide 5 ECE Department
 Alternate DIT FFT structures (continued)

n DIT structure with input natural, output bit-reversed (OSYP

9.15):

Carnegie
Mellon Slide 6 ECE Department
 Alternate DIT FFT structures (continued)

n DIT structure with both input and output in natural order (OSYP
9.16):

Carnegie
Mellon Slide 7 ECE Department
 Alternate DIT FFT structures (continued)

n DIT structure with same structure for each stage (OSYP 9.17):

Carnegie
Mellon Slide 8 ECE Department
 Comments on alternate FFT structures

n A method to avoid bit-reversal in filtering operations is:

– Compute forward transform using natural input, bit-reversed output (as
in OSB 9.10)
– Multiply DFT coefficients of input and filter response (both in bit-
reversed order)
– Compute inverse transform of product using bit-reversed input and
natural output (as in OSB 9/14)

n Latter two topologies (as in OSYP 9.16 and 9.17) are now rarely
used

Carnegie
Mellon Slide 9 ECE Department
 The decimation-in-frequency (DIF)
FFT algorithm

n Introduction: Decimation in frequency is an alternate way of

developing the FFT algorithm
n It is different from decimation in time in its development,
although it leads to a very similar structure

Carnegie
Mellon Slide 10 ECE Department
 The decimation in frequency FFT (continued)

n Consider the original DFT equation ….

N −1
X[k] = ∑ x[n]WNnk
n=0

n Separate the first half and the second half of time samples:
( N / 2)−1 N −1
X[k] = ∑ x[n]WNnk + ∑ x[n]WNnk
n=0 n= N / 2
( N / 2)−1 ( N / 2)−1
= ∑ x[n]WNnk + WN( N / 2)k ∑ x[n + ( N / 2)]WNnk
n=0 n=0
( N / 2)−1
= ∑
n=0
[ x[n] + (−1)k x[n + ( N / 2)]]WNnk
n Note that these are not N/2-point DFTs
Carnegie
Mellon Slide 11 ECE Department
 Continuing with decimation in frequency ...

(N / 2)−1
X[k] = ∑
n=0
[ ]
x[n] + (−1) k x[n + ( N / 2)] WNnk

n For k even, let k = 2r

(N / 2)−1 ( N / 2)−1
X[k] = ∑
n=0
[ x[n] + (−1) 2r x[n + ( N / 2)] WNn2r =] n=0
∑ [ x[n] + x[n + ( N / 2)]]WNnr/ 2
n For k odd, let k = 2r + 1
( N / 2)−1
X[k] = ∑
n=0
[ ]
x[n] + (−1)2r (−1)x[n + (N / 2)] WNn(2r+1)

( N / 2)−1
= ∑ [ x[n] − x[n + (N / 2)]]WNn WNnr/ 2
n=0
n These expressions are the N/2-point DFTs of
x[n] + x[n + ( N / 2)] and [x[n] − x[n + ( N / 2)]]WNn
Carnegie
Mellon Slide 12 ECE Department
 These equations describe the following
structure:

Carnegie
Mellon Slide 13 ECE Department
 Continuing by decomposing the odd and even
output points we obtain …

Carnegie
Mellon Slide 14 ECE Department
 … and replacing the N/4-point DFTs by
butterflys we obtain

Carnegie
Mellon Slide 15 ECE Department
 The DIF FFT is the transpose of the DIT FFT

n To obtain flowgraph transposes:

– Reverse direction of flowgraph arrows
– Interchange input(s) and output(s)

n DIT butterfly: DIF butterfly:

n Comment:
– We will revisit transposed forms again in our discussion of filter
implementation
Carnegie
Mellon Slide 16 ECE Department
 The DIF FFT is the transpose of the DIT FFT

n Comparing DIT and DIF structures:

DIT FFT structure: DIF FFT structure:

n Alternate forms for DIF FFTs are similar to those of DIT FFTs
Carnegie
Mellon Slide 17 ECE Department
 Alternate DIF FFT structures

n DIF structure with input natural, output bit-reversed (OSYP

9.22):

Carnegie
Mellon Slide 18 ECE Department
 Alternate DIF FFT structures (continued)

n DIF structure with input bit-reversed, output natural (OSB 9.22):

Carnegie
Mellon Slide 19 ECE Department
 Alternate DIF FFT structures (continued)

n DIF structure with both input and output natural (OSYP 9.24):

Carnegie
Mellon Slide 20 ECE Department
 Alternate DIF FFT structures (continued)

n DIF structure with same structure for each stage (OSYP 9.25):

Carnegie
Mellon Slide 21 ECE Department
 FFT structures for other DFT sizes

n Can we do anything when the DFT size N is not an integer

power of 2 (the non-radix 2 case)?
n Yes! Consider a value of N that is not a power of 2, but that
still is highly factorable …
Let N = p1 p2 p3 p4 ... pν ; q1 = N / p1, q2 = N / p1 p2 , etc.
n Then let
N −1
X[k] = ∑ x[n]WNnk
n=0
q1 −1 q1 −1 q1 −1
p1rk ( p1r+1)k ( p r +2)k
= ∑ x[ p1r]WN + ∑ x[ p1r + 1]WN + ∑ x[ p1r + 2]WN 1 + ...
r=0 r=0 r=0

Carnegie
Mellon Slide 22 ECE Department
 Non-radix 2 FFTs (continued)

n An arbitrary term of the sum on the previous panel is

q1 −1
∑ x[ p1r + l]WN( p1r+l)k
r=0
q1 −1 q1 −1
p rk
= ∑ x[ p1r + l]WN1 WNlk = WNlk ∑ x[ p1r + l]Wqrk
1
r=0 r=0

n This is, of course, a DFT of size q1 of points spaced by p1

Carnegie
Mellon Slide 23 ECE Department
 Non-radix 2 FFTs (continued)

n In general, for the first decomposition we use

p1 −1 q1 −1
X[k] = ∑ WNlk ∑ x[ p1r + l]Wqrk
1
l =0 r=0

n Comments:
– This procedure can be repeated for subsequent factors of N
– The amount of computational savings depends on the extent to which N
is “composite”, able to be factored into small integers
– Generally the smallest factors possible used, with the exception of some
use of radix-4 and radix-8 FFTs

Carnegie
Mellon Slide 24 ECE Department
 An example …. The 6-point DIT FFT
1 2
n P1 = 2; P2 = 3; X[k] = ∑ W6lk ∑ x[2r + l]W3rk
l=0 r =0

Carnegie
Mellon Slide 25 ECE Department
 Summary

n This morning we considered a number of alternative ways of

computing the FFT:
– Alternate implementation structures
– The decimation-in-frequency structure
– FFTs for sizes that are non-integer powers of 2
– Using standard FFT structures for inverse FFTs

n Starting on Monday we will begin to discuss digital filter

implementation structures

Carnegie
Mellon Slide 26 ECE Department