Gn3311521155

D.Venkata Kishore, C.Ram Kumar / International Journal of Engineering Research and
Applications (IJERA) ISSN: 2248-9622 www.ijera.com
Vol. 3, Issue 3, May-Jun 2013, pp.1152-1155
1152 | P a g e
Design and Implementation of Pipelined FFT Processor
*D.Venkata Kishore, **C.Ram Kumar
*Lecturer, ECE Department, JNTUA College of Engineering Pulivendula
**Lecturer, ECE Department, JNTUA College of Engineering Pulivendula
Abstract
It is important to develop a high-
performance FFT processor to meet the
requirements of real time and low cost in many
different systems. So a radix-2 pipelined FFT
processor based on Field Programmable Gate
Array (FPGA) for Wireless Local Area Networks
(WLAN) is proposed. Unlike being stored in the
traditional ROM, the twiddle factors in our
pipelined FFT processor can be accessed directly.
A novel simple address mapping scheme is also
proposed. The FFT processor has two pipelines,
one is in the execution of complex multiplication
of the butterfly unit, and the other is between the
RAM modules, which read input data, store
temporary variables of butterfly unit and output
the final results. Finally, the pipelined 64-point
FFT processor can be completely implemented
within only 67 clock cycles.
Keywords-FFT; FPGA; address mapping
I. INTRODUCTION
Fast Fourier Transform (FFT) processor is
widely used in different applications, such as
WLAN, image process, spectrum measurements,
radar and multimedia communication services [1].
However, the FFT algorithm is a demanding task
and it must be precisely designed to get an efficient
implementation. If the FFT processor is made
flexible and fast enough, a portable device equipped
with wireless transmission system is feasible.
Therefore, an efficient FFT processor is required for
real-time operations [2] and designing a fast FFT
processor is a matter of great significance.
In the past twenty years, FPGA has
developed rapidly and gradually become universal.
Compared with design flow of traditional ASIC,
designs based on FPGA have the advantages of
flexibility and high performance price ratio. Many
researchers have studied on pipelined FFT based on
FPGA [3], [4], [5]. For instance, in [3], they
proposed an approach to design an FFT processor
for wireless applications, but his design has too
many clock cycles and isn’t fast enough. In
comparison to their designs, we propose a simple
and feasible pipelined implementation of a 32-bit
64-point FFT processor based on FPGA for WLAN.
This paper is organized as follows. In the
next section, we introduce a basic radix-2 FFT
algorithm to briefly discuss which decimation is
better to the system, a three-multiplication method,
and a novel address mapping scheme which reduces
delay and increases the speed of the system. In
section III, the pipelined FFT architecture is
proposed and each unit is also illustrated. Section IV
is the implementation of the 64-point FFT processor
based on FPGA, and hardware resources are
explicitly listed out. The last section gives the
conclusion.
II. FFT ALGORITHM AND ADDRESS
MAPPING SCHEME
A. Radix-2 FFT Algorithm
The FFT algorithm can compute the
Discrete Fourier Transform (DFT) effectively.
Given a sequence {x(n)} of N complex numbers, we
can compute its DFT, another sequence {X(k)} of N
complex numbers, according to the following
formula [6]
N-1
N
X(k) = ∑ x(n)W nk
, k= 0,1,……,N-1. (1)
n = 0
And according to the different way to
decimate, it can be divided into two types, DIF
(Decimation in Frequency) and DIT (Decimation in
Time). The DIF algorithm is easier to design than
DIT. And considering the finite word length effect,
DIF has much more advantages than DIT, such as
reducing the additive noise, which is introduced by
the multiplication when it is implemented with the
fixed point [7] and reducing the complexity of the
whole system. Consequently, we use the DIF
algorithm to design radix-2 FFT module and most of
current FFT processors are also based on this
algorithm [8].
B. Three-multiplication Method
It’s undeniable that complex multiplication
is the dominant factor affecting the speed and the
throughput of FFT processor. Computing a complex
multiplication requires four real multipliers and two
real adders. As we all know, the hardware area of a
real multiplier is larger than that of a real adder in
FPGA. So we should do our best to convert the
complex multiplication into addition and subtraction
to optimize the whole performance as high as
possible. Having taken into account all operands are
32-bit complex numbers, the difference of two

1153 | P a g e
inputs X m (i) and X m ( j) can be
expressed by z1 = x1 + jy1 and the twiddle
factor WN
ek
= exp(2n π)can be expressed by
N
z 2 = x2 + jy 2 . So the product of them can be also
expressed by z = x + jy . That’s to say, the 16-bit real
part x of the product is equivalent to x1 x2 − y1 y2
and the 16-bit imaginary part y is equivalent to x1 y
2 + x2 y1 .
Therefore, we can transform the product z
easily as the following equations
x = x1 ( x 2 + y 2 ) − y 2 ( x1 + y1 ), (2)
y = x1 ( x 2 + y 2 ) − x 2 ( x1 − y1 ). (3)
Obviously, using this factorization scheme,
the system has some advantages [9]. The number of
real multiplications is reduced from four to three.
And addition has less consumption than
multiplication. So the system power consumption is
also reduced. In this study, we can save sixteen
embedded multiplier 9-bit elements in FPGA. As for
this 64-point FFT processor, the numerical values of
x2 + y2 , x1 − y1 , x1 + y1 , x2 and y2 can be gotten
before they participate in the real multiplication.
C. The Novel Address Mapping Scheme
In this paper, the block size of our system
is 64 points. Having considered the properties of
radix-2 64-point FFT, it needs to read 8 operands
from memories at a time so as to achieve a high-
speed FFT. As we all know, parallel accessing data
is crucial to a system [10]. Thus, these 8 operands in
our design are located in different row or column of
memory blocks and this arrangement ensures that 8
conflict-free memory accesses can be performed in
parallel. Initially, we use 8 32-bit dual-port
memories to store 64 operands in sequence. And
then a new linear shift conflict-free address mapping
scheme is adopted to change the addresses of
operands. The primary two-dimensional addresses
of operands will be mapped to new ones.
For instance, we assume that the original
two-dimensional coordinate is (a, b), in which a and
b represent the address of the data in one memory
and the number of the 8 memories, respectively.
Then, we obtain a new conflict-free address (A, B)
by means of the following equations
A=b, (4)
B=(a+b)%8. (5)
In our design, it can be ensured that no
memory location is read from or written to at the
same time, and this new mapping scheme is
feasible, effective and simple.
III. THE PIPELINED FFT PROCESSOR
ARCHITECTURE
For high throughput systems, pipelined
architecture is a good choice, and it is also an ideal
method to implement high-speed long-size FFT
owing to its regular structure and simple control.
The performance of pipelined FFT processor can be
improved by optimizing the structure and saving
hardware resources. The block diagram of our
proposed FFT processor is illustrated in Fig.1. It
consists of four essential units. Control unit, the
kernel of the FFT processor, harmonizes the whole
system. Butterfly unit (BU), which has three-stage
pipelined structure, carries out the complex
multiplication. Two dual-port RAMs are used to
store and output data. And AGU, the abbreviation
Fig 1. The pipelined FFT architecture.
of address generator unit, produces 8 3-bit read
addresses and write addresses.
A. Control Unit
Control unit, which generates all control
signals for the whole system, is responsible for
operation control of the processor. A 48-bit signal
w_con controls the whole FFT processor. And this
signal w_con generates two parameters, write_en
and read_en, to control AGU. It also generates sel1
and sel2 signals to select data from two RAMs, each
of which is made up of 8 32-bit registers. The BU
and the remaining parts are controlled by w_con as
well. This control unit harmonizes all steps of the
FFT processor based on a 7-bit counter.
B. The DIF Butterfly Unit
For FFT algorithm, the central component
is the BU that calculates the sum and difference of
two input data, and plays an extremely important
role in computing the product of the difference and
twiddle factors.

1154 | P a g e
We only use 11 factors W0
64 , W1
64 , W2
64 , W3
64 ,
W4
64, W5
64 , W6
64 , W7
64 , W8
64 , W16
64 , W24
64 to
express all 32 factors that we need in this FFT
processor owing to the nk fact that the twiddle factor
W N can be separated into two components. For
instance, W24
64 can be derived from the product of
W4
64 and W24
64 . Moreover, the value of W0
64 is a
constant 1, and for W16
64 , only a negative sign is
needed to add to the real part of the relevant data,
and then to inverse the real part and imaginary part.
So we can eliminate these two factors, that’s to say,
actually we just need 9 twiddle factors. In addition,
we use 16-bit fixed point decimal to express these 9
twiddle factors. Although the fixed point decimal
arithmetic isn’t precise enough, it can satisfy the
requirements of general systems.
Fig 2. Block diagram of BU.
Due to the fact that the multiplication of
twiddle factors and corresponding data is very
important, three-stage pipeline structure is used
for the complex multiplication to obtain a high
speed computation. The architecture of BU is shown
in Fig. 2.
In the BU, both of the complex inputs are
32 bits, including 16-bit real part and 16-bit
imaginary part. The sum of them needs to be scaled
down by a factor of 2 to avoid arithmetic overflow,
and the same operation is applied to the difference
of them. On the other hand, the factor W0
64 and the
first parameter con_s of the second multiplexer are
not involved in the complex multiplier, and they
can be used as a constant 1, just as Fig.2 has
depicted. Thus, the power consumption of the
complex multiplier can be reduced and the hardware
resources will be saved. There’re some points to be
emphasized. The difference and twiddle factors are
both 32 bits, so the result of the first complex
multiplier will be 64 bits. But because we adopt the
fixed point decimal computation, we should
intercept it to a 32-bit parameter f_mult as the input
of the second complex multiplier.
The most remarkable advantage in this unit
is that we use 3 32-bit registers to realize the three-
stage pipeline of butterfly transform, using register1
to store the difference, register2 to store the
intercepted result of second stage f_mult and
register3 to store the final result cmul_b.
C. The RAM Unit
The RAM1 and RAM2 are made up of 8
32-bit registers respectively. And data is always
written to the outside memories from RAM2, and it
is always read to RAM1 from the outside memories.
Then let introduce the key algorithm used in this
unit. Considering the properties of 64-point FFT, we
can use the radix-2 DIF 8-point FFT as a whole unit,
so there
Fig 3 The example of radix-2 8-point FFT.
are only two stages to accomplish the 64-point FFT.
And these two stages are identical to the six stages
of the standard radix-2 DIF 64-point FFT. System
parallel reads 8 32-bit operands from outer
memories to RAM1 at a time, and we need only
read sixteen times. An example of this algorithm is
shown in Fig.3.
The pipeline in RAM units is briefly
discussed as follows. Firstly, system reads 8
operands from outer memories and writes them to
RAM1, and then the results will be stored in RAM2
after they are computed. Meanwhile, system reads
the next 8 operands. Subsequently, system will
access operands from RAM1 or RAM2 to compute
these 16 operands. Furthermore, within a single
clock four butterfly computations will be executed
simultaneously. The final results of these 16
operands will be all stored in RAM2, and then they
are written to outer memories. At the same time,
another 8 operands will be read. Accordingly, only
76 clock cycles are needed to complete this radix-2
DIF 64-point FFT.
D. AGU
Compared with other units, AGU is also
quite important. It will create 8 read and 8 write
addresses, which determine the data access to outer
memories.

1155 | P a g e
In this FFT processor, we adopt the in-
place computation method so as to make it a more
simplified and faster system. That’s to say, we write
the results into where they are read. In contrast to
the sequence of 64 input operands configured by
address mapping, the final output sequence of this
FFT processor are in bit-reversed order and need to
be adjusted to normal order. And we can make these
appropriate adjustments before the FFT
computation. So, although we adjust the sequence of
inputs or outputs, the performance of our FFT
processor won’t be degraded.
IV. HARDWARE RESOURCES
The functional simulation and timing
simulation are successfully made. The main
hardware resources of this design are given as
follows. The device is the EP2C70F896C6 of
Cyclone II family. The total logic elements are
562/68,416 (8%), the total pins are 563/622(91%)
and the total embedded multiplier 9-bit elements are
48/300(16%). Meanwhile, the clock frequency is
31.69MHz. As in [3], they proposed a fixed-point
l6-bit 64-point FFT processor with 92 clock cycles
in total, but our clock cycles is 67. And our FFT
processor has a higher speed and lower power
consumption.
V. CONCLUSION
This paper proposes a novel radix-2 FFT
processor based on FPGA for WLAN, using Verilog
HDL as hardware description language and Quartus
II as design and synthesis tool. To achieve high-
throughput, pipelined architectures have been used
in the butterfly unit and the dual-port RAM. The
dedicated parallel-pipelined FFT processor
architecture can process input data at high speed,
and the whole system performance can be greatly
improved due to adopting a novel simple address
mapping scheme. For radix-two system, this
mapping scheme is better and simpler than most of
others. The design is implemented on a FPGA chip.
And this pipelined FFT completes a complex 64-
point FFT within 2.1μs. The hardware testing result
explains that it can meet the requirements of the
WLAN.
REFERENCES
[1] J. A. C. Bingham, “Multicarrier modulation
for data transmission: an idea whose
time has come,” IEEE Communication
Magazine, vol. 28, no. 5, pp. 5-14, May
1990.
[2] J. Palicot and C. Roland, “FFT: a basic
function for a reconfigurable receiver,”
10th International Conference on
Telecommunications, vol. 1, pp. 898-902,
March 2003.
[3] Min Jiang, Bing Yang, Yiling Fu, et al.,
“Design Of FFT processor with Low
Power complex mutliplier for OFDM-
based high-speed wireless applications,”
International Symposium on
Communications and Information
Technology, vol. 2, pp. 639-641, Oct.
2004.
[4] Kai Zhong, Hui He, and Guangxi Zhu, “An
ultra-high speed FFT processor,”
International Symposium on Signals,
Circuits and Systems, vol. 1, pp. 37 -
40, July 2003.
[5] Hongjiang He and Hui Guo, “The
Realization of FFT Algorithm based on
FPGA Co-processor,” Second International
Symposium on Intelligent
Information Technology Application, vol.
3, pp. 239-243, Dec. 2008.
[6] J. G. Proakis and D. G. Manolakis,
“Introduction to Digital Signal Processing,”
New York: Macmillan, 1988.
[7] R. B. Perlow and T. C. Denk, “Finite word
length design for VLSI FFT processors,”
Conference Record of the Thirty-Fifth
Asilomar Conference on Signals, Systems
and Computers, vol. 2, pp. 1227– 1231,
Nov. 2001.
[8] J. W. Cooky and J. W. Tukey, “An
algorithm for the machine
calculation of complex Fourier series,”
Math. of Comp., vol. 19, No. 90, pp.
297-301, April 1965.
[9] S. Oraintara, Y. J. Chen, and T. Q. Nguyen,
“Integer fast Fourier transform,” IEEE
Trans. Acoustics, Speech, Signal
Processing, vol. 50, No. 3, pp. 607-618,
March 2002.
[10] J. H. Takala, T. S. Jarvinen, and H. T.
Sorokin, "A Conflict-free parallel memory
access scheme for FFT processors,"
International Symposium on Circuits and
Systems, vol. 4, pp. 524-527, May
2003

Gn3311521155

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Gn3311521155 (20)

Recently uploaded (20)

Gn3311521155