The 3rd IEEE International Symposium on Wireless Systems within the Conferences on Intelligent Data Acquisition and Advanced

Computing Systems
26-27 September 2016, Offenburg, Germany

A Novel Split Radix Fast Fourier Transform Design

for an Adaptive and Scalable Implementation
Adnan Rauf, Muhammad Adeel Pasha and Shahid Masud
Department of Electrical Engineering, SBA School of Science Engineering,
Lahore University of Management Sciences, Lahore, Pakistan

Abstract—Fast Fourier Transform (FFT) is a rudimen- as Fast Fourier Transform exploits the symmetry in
tary operation in signal processing. Modern high speed DFT calculations to speed up the operation. The FFT
signal processing application technologies, such as 4G algorithm which was first implemented by Cooley-Tukey
LTE, 5G, Internet of Things (IoT), etc., necessitate a high exploited the symmetry in DFT operation to reduce the
throughput implementation of FFT. Moreover the FFT size
complexity of arithmetic operations from the order of
required in aforementioned technologies tends to vary with
the operation mode. It is highly desirable to have a FFT
O(N 2 ) to O(N logN ) arithmetic operations [1] .
implementation that not only meets the high throughput Since then Fast Fourier Transform (FFT) has become
demand but is also scalable to cater selectable N-point omnipresent in wide-ranging engineering applications.
FFT. In this paper, our contribution is two-fold; we first It is extensively used in spectrum analysis, high speed
propose a novel split-radix 4/8 FFT algorithm which is image processing, speech recognition, data compression,
comparable in efficiency with radix-8 algorithm but is radar processing, OFDM systems and many more ap-
considerably less complex in design for large size FFTs. plications. High speed application technologies like 4G
Secondly, we developed an automation tool that gives high LTE, 5G communication systems and Internet of Things
level implementation details of an adaptive N-point split-
(IOT) are all the rage in the world at the moment.
radix FFT processor.
These high speed applications require high throughput
Index Terms—DFT, FFT, IoT, split-radix, FFT butterfly, FFT for their operation. The FFT size required for the
parametrized design implementation of these high speed applications also
varies. The FFT size for 3G, 4G LTE, WiMax commu-
I. I NTRODUCTION nication systems is dependent on the operating mode of
Discrete Fourier Transform (DFT) is used to represent the system and it ranges from 64-Point to 2048-Point [2].
a discrete time signal by a series of sinusoidal functions. Similarly, 5G technology is expected to be introduced in
For a complex valued discrete-time sequence x[n] of N 2020. The FFT size for 5G communication is expected
points, the DFT is defined by eq. 1. to be even greater as compared to 4G LTE technology.
In view of the variable FFT size, it is high desirable
N −1 to have a parametrized and adaptive FFT design that

X[k] = (x[n]WNkn k = 0, 1, 2, ..., N ) (1) can be scaled depending upon the operating requirements
n=0 without going through the laborious effort of designing
for each N-Point FFT size configuration.
Where the phase factor is given by eq. 2
The basic FFT design is based on Radix-2 [3] butterfly
block, which was proposed by Cooley-Tukey [1]. For
WNkn = e−j2knπ/N (2)
more efficient implementation, higher radix and Split
The function X[k] represents the frequency domain radix algorithms such as radix-4 [4], radix-8 [5] and
value of the input function x[n] in the discrete time split-radix 2/4 [6], [7], etc. have been employed. All
domain. For direct evaluation of a single point DFT, N conventional FFT algorithms decompose a size N FFT
complex multiplications and (N − 1) complex additions into two parts, namely odd half and even half, and effec-
are required. Consequently, an N-point DFT requires tively reduce the number of multiplications by using the
N 2 complex multiplications and N × (N − 1) complex symmetric properties of the FFT. On the other hand, the
additions. As a result, the computational complexity split-radix algorithms employ a combination of different
of DFT becomes O(N 2 ). For increased number of N radix FFT for decomposition. A split-radix algorithm
points, direct computation of DFT becomes computa- exploits the best of both radix algorithms in terms of
tionally too-intensive. A collection of algorithms known points catered and stages required for FFT computation

978-1-5090-4317-0/16/$31.00 ©2016 IEEE 116

at the expense of non-uniform butterfly structure [6], [7],
[8]. There are two ways to decompose the FFT operation, x[n] x[2r]
namely Decimation in Time (DIT) and Decimation in
Frequency (DIF). If the decomposition is done by divid-
ing the input sequence x[n] into smaller sub-sequences
it is known DIT and if the decomposition is done by
decomposing the output sequence X[k] into smaller sub- x[n+N/2] x[2r+1]
sequences it is known DIF.
-1 WNr
The low order radix algorithms such as radix-2, radix-
4, split-radix 2/4 require a lot of recursive decomposi- Fig. 1. A basic Radix-2 FFT butterfly structure[3]
tions for large N-point FFT computation which results Figure.1 Radix-2 FFT Butterfly [3]
in considerably larger number of steps (stages). As a
result, they fail to meet the throughput demand of high points that are in the power of 4 i.e. (N = 4n ). The
speed applications. On the other hand, higher radix number of stages required for N -point FFT is log4 N
algorithms, such as radix-8, become too complex for where each stage requires N/4 butterflies. The butterfly
large N-point FFT implementation. In order to meet the structure of radix-4 is much more complex than radix-2
FFT throughput demands of high speed applications and butterfly but lesser number of stages are required for a
reduce structural complexity, we are proposing a new N -point FFT operation, hence resulting in a higher speed
split-radix 4/8 algorithm. The proposed algorithm has for more complex applications. A Radix-4, butterfly is
efficiency comparable to radix-8 algorithm but has less shown in figure 2.
structural complexity. Since base of the proposed radix
is 4, it can cater points (N = 4n ) which is another added
advantage over radix-8 algorithm that works in powers x[n] x[4r]
of 8 only.
For an adaptive and scalable implementation of FFT, x[n+N/4] Wn x[4r+1]
we developed a high level design automation tool. The
W2n
tool generates a high-level split-radix 4/8 FFT structure x[n+N/2] x[4r+2]
for selectable N-point FFT implementations. Once the W 3n

structure is built, the tool computes the arithmetic com-

x[n+3N/4] x[4r+3]
plexity of the proposed split-radix 4/8 FFT operation and
that helps us comparing it with other radices i.e. radix-2,
radix-4 and radix-8 FFT operations. Fig. 2. The template of a Radix-4 FFT butterfly [4]
The rest of the paper organized such that in Section
II related work regarding FFT and radix algorithms is Further speedup was achieved by increasing the radix
presented. In Section III, our proposed split-radix 4/8 of FFT operation. A radix-8 [5] decomposes the FFT into
algorithm and automation tool are explained in detail. We eight FFTs recursively. The algorithm can handle points
will represent the qualitative analysis results in Section (N = 8n ) and the number of stages required for N -
IV and finally the paper is concluded in Section V. point FFT is log8 N where each stage has N/8 butterflies.
Again, the butterfly structure of radix-8 is more complex
II. R ELATED W ORK than radix-2 and 4 butterflies but less number of stages
The first implementation of FFT was presented by required for a particular N-point FFT operation. The
Cooley-Tukey in 1965 [1]. Cooley-Tukey exploited the basic radix-8 butterfly is shown in figure 3.
symmetry in DFT to reduce the arithmetic complexity. It was further observed that by combining radix-2
Their work decomposed an N -point sequence into N/2 and radix-4 algorithms, the arithmetic complexity was
even and odd sub-sequences. The algorithm came to greatly reduced. In split-radix 2/4 algorithm [6], [9],
be known as radix-2 algorithm. The number of stages [10], the even output indexes were decomposed based
required for N -point FFT is log2 N and each stage on radix-2 decomposition and odd output indexes were
consists of N/2 butterflies. Basic Butterfly of radix2 DIF decomposed on the basis of radix-4 decomposition. The
FFT [3] is shown in figure 1. resultant algorithm used log4 N number of stages to
To speed up FFT operation radix-4 [4] algorithm was compute N -Point FFT. The drawback of the algorithm
introduced. The algorithm decomposes the FFT into four is the non-uniformity of the structure. The algorithm
FFTs recursively. It operates on base 4 and can handle also uses radix-2 DIF butterflies at the last stage for

117
 x[n] + + + X[8r]

x[n+N/8] + + - * X[8r+1]
Wn

x[n+N/4] + - + * X[8r+2]
W2n
x[n+3N/8] + - * - * X[8r+3]
-i W3n

x[n+N/2] - + + * X[8r+4]
W4n
x[n+5N/8] - * + - * X[8r+5]
(1-i)/2 W5n
x[n+6N/4] - * - + * X[8r+6]
-i W6n
x[n+7N/8] - * - *-i - * X[8r+7]
(-1-i)/2 W7n

Fig. 3. Radix-8 FFT butterfly structure [5]

signal reordering. The split-radix 2/4 butterfly is shown FFT using radix-4 and radix-8 decomposition algorithms.
in figure 4. The proposed split-radix-4/8 algorithm decomposes a
N -point DFT sequence recursively into two N/4 DFT
x[4r] sub-sequences and four N/8 DFT sub-sequences. The
x[n]
Decimation in Frequency (DIF) equations for proposed
radix algorithm are given in figure 5.
x[n+N/4] x[4r+1]
The even indexed outputs are decomposed based on
Wn
radix-4 algorithm and the odd indexed outputs are de-
x[n+N/2] x[4r+2] composed based on radix-8 algorithm. The proposed
i W3n butterfly is shown in figure 6.
x[n+3N/4] x[4r+3]

Stage1 Stage2
Fig. 4. Split Radix-2/4 FFT Butterfly
Figure.4 Radix-2/4 structure
[9] FFT [9]
Butterfly x(0) X(0)
x(1) X(8)
More split-radix algorithms also exist such as split- Radix 4 X(4)
radix 2/8 and split-radix 2/4/8 algorithms [8]. The split- x(2)
X(12)
radix 2/8 uses radix-2 decomposition for even outputs x(3)
and radix-8 decomposition for odd outputs. It reduced
the arithmetic complexity further but the structure be- Radix 4
comes extremely non-uniform due to the presence of
not only radix-2 and radix-8 butterflies but other radix Radix 2
butterflies. In [8], in order to produce a more uniform
structure a split-radix 2/4/8 has been developed based on Radix 2
split-radix 2/8. x(12) Radix 4/8 X(3)
x(13) Radix 2
III. P ROPOSED A PPROACH X(11)
x(14) X(7)
For larger FFTs having more points, the lower radix x(15) Radix 2
X(15)
FFT algorithms incur a huge latency cost as the FFT
computation requires a lot of stages to complete the large
FFT sizes. As a result, these FFT algorithms fail to meet Fig. 7. 16-Point Radix-4/8 FFT Architecture
the throughput demand of high speed applications. On
the other hand, the higher order radix algorithms are The proposed split radix-4/8 algorithm operates on
quite capable for computing FFT efficiently but their N = 4n point sequences and takes log8 N stages to
structure becomes too complex for increased FFT size. compute the FFT of the sequence. As a result, we get
To bridge this gap, we are proposing a new split an algorithm which is more efficient (in number of
radix-4/8 FFT algorithm which decomposes an N -Point stages) than radix-4 FFT algorithm and less complex as

118
 ே ே ଷே ܰ Ȁͺ
ಿ
ሺ ሻିଵ ቂቀ‫ݔ‬ሾ݊ሿ െ ‫ ݔ‬ቂ݊ ൅ ଶ ቃቁ െ ݆ ቀ‫ ݔ‬ቂ݊ ൅ ସ ቃ െ ‫ ݔ‬ቂ݊ ൅ ସ
ቃቁቃ ൅ ܹܰ
Xሾͺ‫ ݎ‬൅ ͳሿ ൌ σ௡ୀ଴
ఴ
ே ହே ଷே ଻ே ͺ݊‫ݎ‬ ݊
ቂቀ‫ ݔ‬ቂ݊ ൅ ଼ ቃ െ ‫ ݔ‬ቂ݊ ൅ ଼
ቃቁ െ ݆ ቀ‫ ݔ‬ቂ݊ ൅ ଼
ቃ െ ‫ ݔ‬ቂ݊ ൅ ଼
ቃቁቃ Ǥ ܹܰ ܹܰ

ே ே ଷே ͵ܰȀͺ
ಿ
ሺ ሻିଵ ቂቀ‫ݔ‬ሾ݊ሿ െ ‫ ݔ‬ቂ݊ ൅ ଶ ቃቁ ൅ ݆ ቀ‫ ݔ‬ቂ݊ ൅ ସ ቃ െ ‫ ݔ‬ቂ݊ ൅ ସ
ቃቁቃ ൅ ܹܰ
Xሾͺ‫ ݎ‬൅ ͵ሿ ൌ σ௡ୀ଴
ఴ
ே ହே ଷே ଻ே ͺ݊‫ݎ‬ ͵݊
ቂቀ‫ ݔ‬ቂ݊ ൅ ଼ ቃ െ ‫ ݔ‬ቂ݊ ൅ ቃቁ ൅ ݆ ቀ‫ ݔ‬ቂ݊ ൅ ቃ െ ‫ ݔ‬ቂ݊ ൅ ቃቁቃ Ǥ ܹܰ ܹ
଼ ଼ ଼ ܰ

ே ே ଷே ܰ Ȁͺ
ಿ
ሺ ሻିଵ ቂቀ‫ݔ‬ሾ݊ሿ െ ‫ ݔ‬ቂ݊ ൅ ଶ ቃቁ െ ݆ ቀ‫ ݔ‬ቂ݊ ൅ ସ ቃ െ ‫ ݔ‬ቂ݊ ൅ ቃቁቃ ൅ ܹܰସ
Xሾͺ‫ ݎ‬൅ ͷሿ ൌ σ௡ୀ଴
ఴ
ே ହே ଷே ଻ேͺ݊‫ݎ‬ ͷ݊
ቂቀ‫ ݔ‬ቂ݊ ൅ ଼ ቃ െ ‫ ݔ‬ቂ݊ ൅ ଼ ቃቁ െ ݆ ቀ‫ ݔ‬ቂ݊ ൅ ଼ ቃ െ ‫ ݔ‬ቂ݊ ൅ ଼ ቃቁቃ Ǥ ܹܰ ܹܰ

ே ே ଷே ͵ܰȀͺ
ಿ
ሺ ሻିଵ ቂቀ‫ݔ‬ሾ݊ሿ െ ‫ ݔ‬ቂ݊ ൅ ଶ ቃቁ ൅ ݆ ቀ‫ ݔ‬ቂ݊ ൅ ସ ቃ െ ‫ ݔ‬ቂ݊ ൅ ସܰ
ቃቁቃ െ ܹ
Xሾͺ‫ ݎ‬൅ ͹ሿ ൌ σ௡ୀ଴
ఴ
ே ହே ଷே ଻ே ͺ݊‫ݎ‬ ͹݊
ቂቀ‫ ݔ‬ቂ݊ ൅ ଼ ቃ െ ‫ ݔ‬ቂ݊ ൅ ଼ ቃቁ ൅ ݆ ቀ‫ ݔ‬ቂ݊ ൅ ଼ ቃ െ ‫ ݔ‬ቂ݊ ൅ ଼ ቃቁቃ Ǥ ܹ ܹ
ܰ ܰ

ಿ
ሺ ሻିଵ ே ே ଷே Ͷ‫݊ݎ‬
XሾͶ‫ݎ‬ሿ ൌ σ௡ୀ଴
ర
ቀሾ‫ ݔ‬ሾ݊ሿ ൅ ‫ ݔ‬ቂ݊ ൅ ସ ቃሿ ൅ ሾሾ‫ ݔ‬ቂ݊ ൅ ଶ ቃ ൅ ‫ ݔ‬ቂ݊ ൅ ସ
ቃሿ ቁ ܹܰ

XሾͶ‫ ݎ‬൅ ʹሿ ൌ σ௡ୀ଴

ሺ ሻିଵ ே ே ଷே ʹܹ݊ Ͷ‫݊ݎ‬
ర
ቀ൤‫ݔ‬ሾ݊ሿ െ ‫ ݔ‬ቂ݊ ൅ ସ ቃ൨ െ ሾሾ‫ ݔ‬ቂ݊ ൅ ଶ ቃ െ ‫ ݔ‬ቂ݊ ൅ ସ
ቃሿ ቁ ܹܰ ܰ

Fig. 5. Decimation in Frequency (DIF) equations for the proposed split Radix-4/8 FFT

compared to radix-8 FFT algorithm. Like in the case of TABLE I

split radix-2/4 algorithm, radix-2 butterflies are required N UMBER OF NON - TRIVIAL COMPLEX MULTIPLICATIONS
at the last stage, our proposed approach requires radix-4
butterflies at the last stage of FFT processing. Multiplications Multiplications
FFT Radix with non-trivial with non-trivial
twiddle factors eight roots of unity
Radix-2 68 30
It has been discovered that as we vary the FFT size Radix-4 56 20
N , the pattern with which the butterflies co-relate in Radix-2/4 52 20
FFT computation also changes. The 16-Point and 64- Radix-8 44 36
Point FFT implementation for proposed radix algorithm Radix-4/8 44 30
are shown in figures 7 and 8. It can also be observed
from the figures that a 16-Point FFT butterfly structure
can be used to construct the next higher order (64- IV. I MPLEMENTATION C OMPLEXITY R ESULTS AND
Point) FFT processor, hence introducing the chances of A NALYSIS OF R ADIX -4/8 FFT
a modular and scalable implementation where a basic
building block can be reused to construct the higher- To compare the proposed FFT implementation with
order or more-complex FFT architectures. the existing ones in the literature, we analyzed the
total number of arithmetic operations required by them.
The number of non-trivial complex multiplications with
In order to develop an adaptive and scalable imple- twiddle factors and eight complex root of unity for 64-
mentation, we designed an automation tool. The automa- Point FFT are given in Table I. In [7], it is stated that a
tion tool, depending upon the selected FFT size, gives complex multiplication consists of three real multiplica-
the high-level implementation details of the FFT oper- tions and three additions. Whereas, multiplication with
ations and forms the split-radix 4/8 butterfly structure complex root of unity consists of two real multiplication
for the required size FFT operation. Whereas the low- and two real additions. Inferring the same relation, the
level implementation details of the butterflies are left to number of non-trivial real additions and multiplications
the designer. The design tool also calculates arithmetic for a 64-Point FFT is given in Table II and we can
complexity of the final split-radix implementation that see that our proposed split radix-4/8 implementation is
can be used to compare it with radix-4 and radix-8 consuming the least number of additions and non-trivial
algorithms. real multiplications that would directly impact the area

119
 x(0) x(0) + x(2) +x(4) +x(6)

x(1)
x(1) + x(3) +x(5) +x(7)

x(0) - x(2) +x(4) -x(6) * W2n

x(2) -
+ - W2n
[x(1) - x(3) +x(5) -x(7)] * -j * W 2n
x(3)
- -j W2n
+ {(x(0)-x(4) -j[x(2)-x(6)]) + (x(1)-x(5) -
x(4) - x(0)-x(4) -j[x(2)-x(6)]
j[x(3)-x(7)]) * WN/8} * Wn

- + Wn
x(0)-x(4) +
(x(1)-x(5) -j[x(3)- {(x(0)-x(4) -j[x(2)-x(6)]) - (x(1)-
x(5) x(7)]) * WN/8 x(5) -j[x(3)-x(7)]) * WN/8} * W5n
N/8
- x(1)-x(5) +
W - W5n
{(x(0)-x(4)+j[x(2)-x(6)]) + ( x(1)-x(5)
x(6) + j[ x(3)-x(7) ])* W3N/8} * W3n
-j - x(0)-x(4)+j[x(2)-x(6)] 3n
- x(2)-x(6) -j[x(2)-x(6)] + W
(x(1)-x(5)+j[x(3)- {(X(0)-x(4)]+j[X(2)-x(6)]) – (X(1)-x(5)
x(7)]) * W3N/8 + j[ X(3)-x(7) ]) * W3N/8} * W7n
x(7) 7n
-j - W3N/8 - W
- x(3)-x(7) -j[x(3)-x(7)]

Fig. 6. Butterfly structure for the proposed split Radix-4/8 FFT

TABLE II
C OMPARATIVE A NALYSIS OF ADDITIONS AND NON - TRIVIAL REAL MULTIPLICATIONS

Multiplications Multiplications Total real Total real

FFT Radix with non-trivial with non-trivial multiplications additions
twiddle factors eight roots of unity
Radix-2 68 30 264 1032
Radix-4 56 20 208 976
Radix-2/4 52 20 196 964
Radix-8 44 36 204 972
Radix-4/8 44 30 192 960

and computation time complexity of the target hardware. R EFERENCES

The same fact is further highlighted by the graph given
[1] J. W. Cooley and J. W. Tukey, “An Algorithm for the Machine
in figure 9. Calculation of Complex Fourier Series,” Mathematics of Com-
putation, vol. 19, no. 90, pp. 297–301, 1965.
V. C ONCLUSION [2] P. P. Boopal, M. Garrido, and O. Gustafsson, “A Reconfigurable
FFT Architecture for Variable-Length and Multi-Streaming
In this paper, we have presented a novel split-radix 4/8 OFDM Standards,” in IEEE International Symposium on Cir-
FFT algorithm that can be used to meet high through- cuits and Systems (ISCAS). IEEE, 2013, pp. 2066–2070.
put demands of modern high throughput applications. [3] A. Haveliya, “Design and Simulation of 32-point FFT using
Radix-2 Algorithm for FPGA implementation,” in Second In-
The proposed algorithm has less arithmetic complexity ternational Conference onAdvanced Computing and Communi-
as compared to its counterparts while maintaining a cation Technologies (ACCT). IEEE, 2012, pp. 167–171.
comparable speed. In order to implement an adaptive [4] A. Ferizi, B. Hoeher, M. Jung, G. Fischer, and A. Koelpin,
and scalable FFT for parametrized N -Point FFT, we “Design and Implementation of a Fixed-Point Radix-4 FFT
Optimized for Local Positioning in Wireless Sensor Networks,”
developed a high level automation and implementation in 9th International Multi-Conference on Systems, Signals and
tool. Depending upon selected N -Point FFT size, the Devices (SSD). IEEE, 2012, pp. 1–4.
tool develops a high-level structure for implementation. [5] M. Sun, L. Tian, and D. Dai, “Radix-8 FFT Processor Design
based on FPGA,” in 5th International Congress on Image and
Once structure is formed the mathematical consistency
Signal Processing (CISP). IEEE, 2012, pp. 1453–1457.
and speed up achieved is calculated that can be used to [6] P. Duharnel and H. Hollmann, “Split Radix FFT Algorithm,”
gauge the arithmetic complexity. Electronics Letters, vol. 20, no. 1, pp. 14 – 16, Jan 1984.

120
 Implementation Complexity Comparison of Radix-4/8 vs. Other Radices
1200
1032
976 964 972 960
1000
800
600
400 264
208 196 204 192
200
0
Radix-2 Radix-4 Radix-2/4 Radix-8 Radix-4/8
Total Real Multiplications Total Real Additions

Fig. 9. Implementation Complexity Comparison of Radix-4/8 vs. Other Radices

[7] H. V. Sorensen, M. T. Heideman, and C. S. Burrus, “On com-

puting the Split-Radix FFT,” IEEE Transactions on Acoustics,
Stage1 Stage2 Stage3 Speech and Signal Processing, vol. 34, no. 1, pp. 152–156,
x(0) X(0) 1986.
x(1) Rad-4 X(32)
x(2) X(16) [8] L. Jia, Y. Gao, J. Isoaho, and H. Tenhunen, “A new VLSI-
x(3) X(48) oriented FFT Algorithm and Implementation,” in 11th Annual
Rad-4 IEEE International ASIC Conference. IEEE, 1998, pp. 337–
341.
Rad-2 [9] N. R. Reddy, L. B. Das, A. Rajesh, and S. Enjapuri, “ASIC
Rad-2 Implementation of High Speed Fast Fourier Transform based
Radix 4/8 Rad-2 on Split-Radix Algorithm,” in International Conference on
Embedded Systems (ICES). IEEE, 2014, pp. 22–27.
Rad-2 [10] Z. Qian and M. Margala, “A novel Low-Power and In-Place
Split-Radix FFT Processor,” in Proceedings of the 24th Great
Rad-4 Lakes Symposium on VLSI (GLVLSI). ACM, 2014, pp. 81–82.

Rad-4
Rad-2
Rad-2
Radix 4/8 Rad-2
Rad-2
Rad-2
Rad-2
Radix 4/8
Rad-2
Rad-2

Radix 4/8 Radix 4/8

Rad-2
Rad-2
Radix 4/8

Rad-2
x(60) Rad-2 X(15)
x(61) X(31)
x(62) Radix 4/8 X(47)
x(63) X(63)

Fig. 8. 64-Point Radix-4/8 FFT Architecture

121