Master’s Thesis Report
Design of FFTs using CORDIC and Parabolic
Synthesis as an alternative to Twiddle
Factor Rotations
By
Muhammad Waqas Shafiq
Nauman Hafeez
Department of Electrical and Information Technology
Faculty of Engineering, LTH, Lund University
SE-221 00 Lund, Sweden
Abstract
Fast Fourier Transform (FFT) processor is an important signal processing block
widely adopted in various disciplines, such as within the computer area, computer
graphics, digital signal processing, communication systems, robotics, navigational
systems, etc. Advanced and complex algorithms in these disciplines need higher
computation performance and low power utilization processes. Due to
advancements and down scaling of hardware technologies, it becomes possible to
fulfill these higher demands from the most advanced algorithms with higher clock
rates on the chips. The improvement in hardware technology performance has
moved the interest more over to the hardware implementation of algorithms.
The scope of this thesis is to investigate different algorithms to compute the
rotations, often done by complex twiddle factor multiplications in in run time, in
FFT processor architectures. The FFT processor is commonly implemented with a
complex multiplier but due to demand of higher point FFTs in the applications, the
size of ROM in the multiplier based implementation for the twiddle factors
becomes the matter of concern with larger chip area. The aim of this thesis is to
design an FFT with un-rolled CORDIC and Parabolic Synthesis algorithms in
order to replace complex multipliers and ROMs.
The outcome of the presented thesis is comparison results for accuracy, power
consumption, performance and area of these algorithms in the STM 65nm CMOS
technology with different transistor technologies and power supplies. It is noted
that implementation with Parabolic synthesis algorithm gives better results as it
automatically gives a high degree of parallelism that gives a shorter critical path
and thereby fast computation. The structure of the methodology will also assure an
area efficient hardware implementation if higher point FFTs are implemented.
2
Acknowledgments
It is a pleasure to thank the many people who made this thesis possible.
It is difficult to overstate our gratitude to our supervisor, Dr. Peter Nilsson. With
his enthusiasm, his inspiration, and his great efforts to explain things clearly and
simply, he helped to make this thesis fun for us. Throughout our thesis work, he
provided encouragement, sound advice, good teaching, good company, and lots of
good ideas. We would have been lost without him. We are also grateful to Erik
Hertz who helped us to understand his research work.
We would like to thank the many people who have taught us throughout the MSc
especially Joachim Rodrigues, Viktor Öwall, Markus Törmänen, Pietro Andreani
and all Teacher assistants for their kind assistance with our courses and lab works
especially Yasser Sherazi, Chenxin Zhang, Reza Meraji, Johan Löfgren, Stefan
Molund, and Jonas Lindstrand.
We wish to thank our best friends Minna, Waqas Khan, Umair, Farhan, Muaz,
Salman, Naveed, Adnan, Shoaib, Ammar and Naeem for providing a stimulating
and fun environment in which I learnt and grow.
We wish to thank our entire families for providing a loving environment for us.
My brothers, my sister Kayynat were particularly very supportive and had faith in
me.
Lastly, and most importantly, we wish to thank our parents. They bore us, raised
us, supported us, taught us, and loved us. To them I dedicate this thesis.
3
Table of Contents
1.
2.
Introduction ................................................................................................. 7
1.1
Overview .............................................................................................. 7
1.2
Organization of Thesis .......................................................................... 8
Fast Fourier Transform ................................................................................. 9
2.1
Discrete Fourier Transform ................................................................... 9
2.2
FFT Algorithms ..................................................................................... 9
2.2.1
Decimation in Time (DIT) algorithm............................................. 10
2.2.2
Decimation in Frequency (DIF) algorithm .................................... 10
2.3
3.
4.
CORDIC ...................................................................................................... 15
3.1
Rotation ............................................................................................. 15
3.2
CORDIC Formulation........................................................................... 18
3.3
Unrolled CORDIC ARCHITECTURE........................................................ 20
3.4
CORDIC Scaling Factor ........................................................................ 22
3.5
Generalize CORDIC ............................................................................. 22
Parabolic Synthesis .................................................................................... 25
4.1
The Methodology ............................................................................... 25
4.1.1
Normalization ............................................................................. 25
4.1.2
Hardware Architecture Development.......................................... 25
4.1.3
Methodology for developing sub-functions ................................. 26
4.2
5.
Hardware Architecture ....................................................................... 12
Hardware Implementation ................................................................. 28
4.2.1
Pre-Processing ............................................................................ 28
4.2.2
Processing .................................................................................. 29
Results ....................................................................................................... 35
5.1
Area ................................................................................................... 36
5.2
Timing ................................................................................................ 37
5.3
Static Power ....................................................................................... 38
4
5.4
Dynamic Power .................................................................................. 40
5.5
Energy ................................................................................................ 41
6.
Conclusions ................................................................................................ 45
7.
Future Work............................................................................................... 46
References......................................................................................................... 47
List of Figures..................................................................................................... 48
List of Tables ...................................................................................................... 49
5
Preface
This thesis is submitted in partial fulfillment of the requirements of Master’s
degree in System on Chip at LTH, Lund University. This thesis has been made
solely by the authors, however most of the text is based on the research of others,
we tried to provide references where necessary.
In this thesis work, initial literature review have been done together with
discussions and note taking to proceed with development process. Nauman Hafeez
has done design of 14 stage un-rolled CORDIC and Parabolic Synthesis design in
VHDL, tested and verified their simulation with MATLAB design. Waqas Shafiq
has done design of 16-point FFT, test and verified with MATLAB results. He has
also done MATLAB reference model design. Result extractions of area, speed and
power after synthesis process have been done together.
6
CHAPTER
1. Introduction
1.1 Overview
In past few years the advancements in VLSI technologies have opened many
windows for designing real time applications, using efficient algorithms with
custom chips. The use of special arithmetic techniques, parallel and pipelined
architectures have led the designs far away from basic computers.
Nowadays the world of information technology has lined up many areas together
such as communication systems, computer graphics, robotics, navigations,
astrophysics etc. and all applications demands high performance with minimum
possible power consumption. Many digital signal processor designers and
manufacturers are facing one big challenge, that is, how to perform complex
mathematical function calculations more efficiently. To gain efficiency, it is
needed to dig inside our complete designing cycle which contains algorithms,
architectures, hardware technology, power supply etc.
Fast Fourier Transform (FFT) is widely used transform in digital applications
especially in communication systems. An FFT is an important processing block in
these systems, which takes most of the hardware complexity in a digital baseband
transceiver for instance. Due to demand in higher data rates in communication
systems, large-point FFTs are required for multiple carrier modulation, such as
1024/2048/8192 etc. An FFT is commonly implemented with complex multiplier;
a complex multiplier is equivalent to four real multipliers and two real adders, and
a ROM to store the twiddle factors. The ROM in this type of implementation takes
most of the chip area, consumes more power and degrades the speed because of
ROM read operation ROM size increased with large-point FFTs. Hence poor
performance of the FFT in terms of power, speed, and area can be seen.
The main objective of the thesis is to investigate algorithms which can be used to
calculate trigonometric functions and perform complex multiplications. An unrolled CORDIC architecture is widely used instead of complex multiplier and
ROM based architectures. Power consumption, speed and accuracy are issues in
this architecture implementation. A newly invented “Parabolic Synthesis”
algorithm is also investigated that could improve the results with respect to
CORDIC based implementation of the FFT. Both algorithms are implemented in
VHDL, synthesized and compared in terms of power, area, accuracy and timing.
Both algorithms are tested to calculate the rotations and to perform complex
multiplications for a 16-Point FFT processor.
7
Traditional methodologies such as use of complex multipliers with pre-computed
twiddling factors (stored in ROM/Registers) for FFT processors are also being
compared with CORDIC and Parabolic Synthesis.
1.2 Organization of Thesis
Chapter 2 explains the basic theory and fundamental equations of CORDIC
algorithm along with the hardware architecture.
Chapter 3 contains the brief description of Parabolic Synthesis and covers the
basic theory, mathematical equations and hardware architectures used to
implement this algorithm for FFT processors.
Chapter 4 is a review of FFT and explanation of butterfly architecture is provided,
and Single-path Delay Feedback (SDF) architecture for implementing FFT is also
discussed and implemented in hardware.
Chapter 5 contains the synthesis results of tested algorithms and gives the
guideline to the reader about the selection of algorithm, the detail comparison
between power, area and speed using different transistor technologies and supply
voltages, is presented.
Chapter 6 summarizes the results achieved in the project.
Chapter 7 discusses the future work in order to improve performance.
8
Chapter 2
2. Fast Fourier Transform
The Fast Fourier Transform (FFT) architecture was invented in 1965 by Cooley
and Tukey [8] and is not a different algorithm from DFT. It is architecture for
efficient computation of DFT.
2.1 Discrete Fourier Transform
DFT is the most widely used transform of all the available transforms in digital
signal processing. The DFT maps the input sequence ( ) into frequency domain.
The Discrete Fourier Transform (DFT) of ( ) input sequence is defined as in
(2.1), which is an N-point sequence [9].
( ) =
( )
= 0,1, … . ,
Where ( ) and ( ) are complex in general and the indices
The term
1
and
(2.1)
are integers.
is twiddle factors, which is described in (2.2)
=
= cos
2
2
(2.2)
The DFT is complex valued but in hardware, the real (Re) and imaginary (Im)
parts are separated in two real parts as in (2.3)
( )=
cos
( )=
sin
sin
+
(2.3)
cos
2.2 FFT Algorithms
Fast Fourier Transform (FFT) [8] is an efficient algorithm for computing the DFT.
Its principle is based on decomposing the computation of the discrete Fourier
transform with sequence of length N into successively smaller discrete Fourier
transforms. Computation complexity is reduced from O(N2) to O(Nlog(N)) after
9
introducing the FFT architecture to calculate a DFT. There are different algorithms
of FFT based on different decomposition schemes. Some of the algorithms are
described next.
2.2.1 Decimation in Time (DIT) algorithm
The algorithm which decomposes the sequence ( ) into subsequently smaller
sequences is called decimation in time algorithm [10]. In order to explain it, a
radix-2 DIT algorithm is explained to demonstrate the decomposition.
In (2.1), the input sequence is divided into even and odd numbered sequences. In
this case, the DFT is described as
( )=
(
)
(
+ 1)
The input sequence ( ) is split into even numbered (
numbered ( + 1) sequence.
= 0,1, . . ,
(2.4)
) sequence and odd
For larger N, the complexity is reduced from O(N2) to nearly half by dividing the
input sequence into even and odd numbered sequences. The complexity can be
further reduced by using the symmetry and periodicity of twiddle factors.
Figure 2.1 A Radix-2 DIT Butterfly
Due to the even and odd numbered sequence split, the twiddle factor is also
simplified. The architecture of radix-2 DIT butterfly only needs one complex
multiplier and two complex adders instead of two complex multipliers.
2.2.2 Decimation in Frequency (DIF) algorithm
In this algorithm, the output of the DFT is divided into smaller subsequent
sequences. DFT main equation is recalled as in (2.1).
Even numbered output sequence can be described in (2.5)
10
(2 ) =
( )
( )
+
,
By changing the second summation from
and by using the property of
(2 ) =
( )+
= 0,1, . . ,
1
(2.5)
to N into the summation from 0 to
, (2.5) can be rewritten as in (2.6)
+
,
= 0,1, . . ,
1
In the same way, odd numbered sequence can be expressed as in (2.7)
(2 + 1) =
( )
+
,
2
= 0,1, . . ,
2
1
(2.6)
(2.7)
(2.6) and (2.7) are combined to get the first order DIF derivation. First step is to
compute ( )
+
and ( )
+
and then multiply the latter
term with
. Fig. 2.2 shows a simple radix-2 DIF butterfly.
in1
+
out1
W
in2
-+
x
out2
Figure 2.2 A Radix-2 DIF Butterfly
A 16-point radix-2 DIF FFT architecture is chosen to be used in this project. The
signal flow graph of the architecture is presented in (2.3).
11
Figure 2.3 A 16-point radix-2 DIF FFT algorithm
Fig. 2.3 shows the signal flow graph for a 16-point FFT. Two structures as the
one in the figure are needed, one for the real part and one for the imaginary part.
The output is in bit-reversal order i.e., the bits are flipped at the output.
The signal flow graph in Fig. 2.4 shows 4 stages of butterflies. In between a
multiplication with the twiddle factor W is shown according to (2.7).
2.3 Hardware Architecture
Figure 2.4 describes a Single-path Delay Feedback (SDF) architecture of a 16point FFT using multipliers. The pre-computed twiddle factors are stored in an
array or ROM if large-point FFTs are to be designed. A controller is used to
generate the addresses to access the stored twiddle factors and to control
multiplexers in every clock cycle.
12
Fixed point 2’s complement data format is used. In 2’s complement, first bit of a
positive number is always 0 and rest of bits are ordinary binary values. For
negative numbers, invert its standard binary value and add one to it.
Complex data (real and imaginary) has been used for both twiddle factors and
input sequence. Both real and imaginary values are 15 bits each; 1 sign bit, 4
integer bits and 10 fractional bits to get 3 decimal places of accuracy in the
fractional part. Data scaling and truncation between different stages of the
architecture due to multiplication is done according to precision requirements.
Fig. 2.4 shows an architecture for a 16-point FFT architecture with multipliers. In
the first phase, 8 input samples are, coming serially, are stored in the 8 registers of
15 bits each during the first stage.
Figure 2.4 A 16-point FFT architecture using multipliers
In the second phase, the first 8 samples, stored in registers, are processed, in the
butterfly, with the remaining 8 samples. . The reason behind is that the 1 st sample,
x(0) has to be processed with the 9th sample, x(8). The second sample, x(1) should
be processed with 10th sample, x(9). The butterfly will generate the sum and
difference of two input samples. The result of difference of input samples will
eventually be stored in the register and multiplied with twiddle factor. In contrast,
the addition of two input samples is sent to the next stage without any
multiplication with a twiddle factor. The MUXes in this architecture are controlled
by a controller, which keep track of the values that are sent to the next stage in the
specific clock cycle. In the second stage, the same operations are done but in two
parts. The first 4 results from the first stage are stored in the 4 registers in the
second stage. These 4 samples are processed with the upcoming 4 results in the
butterfly. The same procedure is done with the next results from the first stage.
This architecture generates the first end result in the 17th clock cycle. It produces
all the 16 results in 32 clock cycles.
13
Twiddling factors approximation and complex multiplication can be done by a
coefficient ROM and a multiplier, by using CORDIC algorithm, and by Parabolic
Synthesis as described in Fig. 2.5.
Figure 2.5 A 16-point FFT architecture
Figure 2.5 shows a complete architecture of a 16-point FFT with both real and
imaginary parts. Complex multiplier (comprising of 4 real multipliers and 2 real
adders) and ROM can be replaced with CORDIC and Parabolic Synthesis
algorithms to calculate real and imaginary values for next stage in the FFT. An
angle “ ” and previous stage real and imaginary values are provided to the
CORDIC algorithm, it will generate the final Xre and Xim values at the output for
the next stage of the FFT. In the same manner, an angle “ ” and previous stage
real and imaginary values are provided to the Parabolic Synthesis algorithm,
which will generate the rotations, with the final Xre and Xim values at the output
according to (2.3).
14
CHAPTER
3. CORDIC
The CORDIC (COordinate Rotation DIgital Computer) is a simple and efficient
“shift-add” algorithm to calculate the wide range of functions including
trigonometric, logarithmic, hyperbolic and linear. It is commonly used when there
is no hardware multiplier is available.
The CORDIC algorithm was first described by Jack E.Volder in 1959. It was
developed to provide the digital solution for the real-time navigation problems [1].
Two basic CORDIC modes are well known for the computation of different
functions, the rotation mode and vector mode. The CORDIC algorithm can be
taken in as iterative structure of additions, subtractions and binary shift operations,
which can perform fixed angle rotation (also known as micro-rotations).
The CORDIC algorithm is well suited for VLSI implementations due to the
simplicity of the involved operations. The CORDIC does not provide the perfect
rotation when it is involved with the complex multiplication because the rotated
vector gets scaled. To achieve perfect angle rotation, the scaling factor can be
easily corrected. The idea behind the CORDIC for rotating the complex number by
successive constant angles is highly useful and needs to be compared with other
algorithms, which are claiming the similar functionality.
3.1 Rotation
As the nature of CORDIC algorithm is based on simplicity, and the basic
observation of unit-length vector can give the foundation of CORDIC algorithm.
The point z in Fig. 3.1 at coordinates (x, y) = (1, 0) is rotated by the angle and
the new point z' will be at (x, y) = (cos ,sin ). Thus, the trigonometric functions
cos and sin can be computed by finding the co-ordinates of point z', which is a
rotated vector of angle [2].
The rotation of the vector in a rectangular coordinate system from one point to
another point can be performed by simply multiplying the vector by a complex
number. Let the initial vector is
z = x + jy
15
(3.1)
Y
r=
1
z’=(cos ,sin )
z=(1,0)
X
Figure 3.1 Unit-length vector in Cartesian coordinate system
And the desire vector is z' which is rotated by angle
z = x + jy
(3.2)
z =ze
(3.3)
z' can easily be computed by multiplying z with
e = cos + sin
z = (x + jy)(cos + sin )
z = x cos + jx sin + jy cos
z = (x cos
(3.4)
(3.5)
y sin
y sin ) + j(x sin + y cos )
(3.6)
(3.7)
Equation (3.7) gives us the new rotated points of the vector z , as shown in Fig.
3.2.
The complex number rotations in 2-dimensions are commutative [3], unlike in
higher level of dimensions. By observing the equation (3.5), whenever two
complex numbers gets multiplied with each other, their phase is added up and the
magnitude gets multiplied. Thus, during the multiplication with the complex
number with unity magnitude we are not affecting the magnitude but only rotating
the vector by the desired angle. Such a rotation is known as a Real rotation as
shown in Fig. 3.2.
16
Figure 3.2 Real Rotation of the vector
x = x cos
y sin
y = x sin + y cos
(3.8)
(3.9)
Let us take (3.8) and (2.9) one step further and make our way towards the
CORDIC
x = cos (x
y sin / cos )
y = cos (y + x sin / cos )
x = cos (x
y tan )
y = cos (y + x tan )
(3.10)
(3.11)
In (3.10) and (3.11) the factor cos provides the scaling in rotated angle and
causes the reduction in magnitude due to that fact that cos
1, but the term
tan brings ease in computation because it is simply a shift operation.
To further investigate the (3.10) and (3.11) and neglecting the scaling factor cos
from the above equation
17
x =x
y tan
y = y + x tan
(3.12)
(3.13)
Equation (3.12) and (3.13) are the coordinates of the rotated vector without any
scaling factor, the absence of scaling factor can affect the results during the
rotations but at the end of total computation this factor can be multiplied to
balance the effect. The new rotated values x and y are shown in Fig. 3.3, such
rotation is known as Pseudo Rotation. Fig. 3.3 shows the different between the real
rotation and pseudo rotation.
Figure 3.3 Real and Pseudo Rotation of the vector
3.2 CORDIC Formulation
Jack E.Volder has proposed the idea to break the rotation angle into the series of
smaller angles i [4] such that the resultant series could utilize the property of
tangent function
tan
=
The resultant series of angles is shown below in Table 3.1
18
(3.14)
Table 3.3.1 Pre-Computed angle set
tan
i
deg
0
45
1
26.5
0.5
2
13.25
0.25
3
7.125
0.125
4
3.576
0.0625
5
1.789
0.03125
6
0.895
0.015625
7
0.447
0.00781
8
0.223
0.00390
9
0.115
0.001953125
10
0.055
0.0009765625
1
Using (3.14) in (3.12) and (3.13), the foundation of the CORDIC equation is
verified as, which is
=
=
+
2
(3.15)
2
(3.16)
The rotations in each stage around the desired angle could be clock-wise or anticlock-wise, and with the sequential values of i the angle
and tan
goes on
decreasing as shown in following equations.
= tan
2
=±
tan
……..
2
tan
2
19
tan
2
tan
2
…….
During the angle rotation, the resultant angles either perform addition or
subtraction depending on the operation deciding factor . Generally, the
CORDIC algorithm for other mathematical functions depends on
as-well.
=
=
+
2
2
(3.17)
(3.18)
Example: To compute the angle = 30º, using the rotation algorithm, is done by
breaking the angles into smaller parts
30º
45º – 26.6º + 14º – 7.1º + 3.6º +1.8º – 0.9º + 0.4º
All the angles in series against the angle 30º can be taken from the Table 3.1.
These angle values can easily be computed using shift, add and subtract
operations.
3.3 Unrolled CORDIC ARCHITECTURE
An eleven stage CORDIC is shown in Fig. 3.4. In the eleven adders at the top, the
remaining angle is computed for each stage. The input variable is the angle for
the cosine and sine that is searched for. At the top, the fixed coefficient angle
values , for each rotation are provided. These are added or subtracted from in
each stage. They are fixed coefficients (hardware wired) but they can be stored in
a ROM as well.
In the middle and the lower adder rows the approximation of the real and
imaginary signals, Xre(n) and Xim(n), are computed and provided to the right. The
initial vector values, x re(1) and xim(1), are provided to the left. The input signals
are taken from previous real and imaginary butterflies (in the case of FFT).
Furthermore, the output signals are provided to next real and imaginary butterflies.
In each stage new vectors are determined, in order to converge towards the vector
that approximates the vector that represents the angle . In each vector rotation
stage there is a crosswise addition or subtraction of the vector coordinates. The
vector rotation depend on the sign bit, sgn, in each stage of the upper row of
adders. That is, the sign bits determine if it should be an addition or subtraction.
There are also divisions of the vector coordinates by a factor corresponding to 2 i
where i’s are the integers {1, 2, 3, 4, 5, 6 … n}. That is, division by {1, 2, 4, 8, 16,
32 … 2n}. This corresponds to the right shifts as discussed in the CORDIC
algorithm, which are done by shifting the busses between the stages. There is thus
no extra cost in hardware for the divisions.
20
Figure 3.4 An un-rolled 14-stage CORDIC architecture
21
3.4 CORDIC Scaling Factor
The above CORDIC algorithm introduces the constant scaling factor (K) with the
resulted rotated vector, and this factor needs to be multiplied. The property
cos
cos(
brings more ease to pre-compute the scaling factor because, the
direction of rotation does not matter in case of cosine function
K=
(1)
….
K = 0.707 x 0.894 x 0.970 x 0.992 x 0.998 ….
K
0.60725
Now, constant scaling factors which can be further designed with shifts and add
operations to eliminate the need of multiplier, as shown in Fig. 3.5.
Figure 3.5 Hardware Multiplication of scaling factor.
3.5 Generalize CORDIC
Generalized CORDIC covers the equations for the linear, circular and hyperbolic
systems and the (3.17) and (3.18) can be modified to accommodate these functions
as well
=
22
2
(3.19)
=
=
+
+
2
(3.20)
(3.21)
The parameters for generalized CORDIC algorithm are listed in Table 3.2 with the
rotation type
Table 3.3.2 Parameters for Generalized CORDIC
Linear Rotation
Circular Rotation
Hyperbolic Rotation
= 2
= tan
= tanh
23
µ=0
2
2
µ=1
µ=-1
24
CHAPTER
4. Parabolic Synthesis
Parabolic synthesis methodology [6] is devised to implement approximation of
different unary functions e.g. trigonometric, logarithm, square root and division
functions that are extremely important in the field of astronomy, digital signal
processing, image processing, robotics, navigation systems etc. High speed
applications in these fields need hardware implementation where software
solutions in most of the cases are not sufficient.
Parabolic synthesis methodology performs an approximation of original unary
function by developing other sub-function and help functions. The architecture of
the proposed methodology itself parallel in nature and take great advantage of
parallelism to reduce the execution time. Only low complexity operations e.g.
multiplications, shifts, additions are used that is easily implemented in the
hardware.
Parabolic synthesis methodology has been devised in order to generate better
results in terms of speed, area, power and precision as compared to algorithms
employed previously for the approximations of unary functions. Pre-processing
and post-processing are two steps besides the most important approximation part
that are used to normalize and transformation respectively. Therefore, the
implementation is divided into three parts, normalizing, approximation and
transforming.
4.1 The Methodology
Pre-processing is the first step; normalization, approximation is the next one in
which sub-function and help functions are developed and post-processing is the
last step to transform the result according to the requirement.
4.1.1 Normalization
Normalization is done to facilitate the hardware implementation by restricting the
numerical range. The purpose of this step is to satisfy that the values should be in
the interval of 0
< 1on the x-axis and 0
< 1on the y-axis keeping the
coordinates of starting point at (0,0) and ending point having coordinates smaller
than (1,1). Furthermore the function must be strictly concave or convex during the
interval.
4.1.2 Hardware Architecture Development
Low complexity operations are used when developing hardware architecture that
approximates an original function. These operations include additions,
multiplications and shifts, which are efficient for hardware implementation. The
emergence of efficient multiplier architecture and downscaling of semiconductor
25
technologies has made the multiplication operation efficient in hardware
implementation.
The proposed methodology is based on decomposition of basic functions. It is
based on second order parabolic functions since these functions can easily be
implemented with low complexity operations. The recombination process is done
with multiplication.
The methodology is devised in terms of second order parabolic functions called
sub-functions ( ), these sub-functions when recombined give the approximation
( ) as shown in (4.1). The accuracy of the results
of the original function
depends on the number of sub-functions used.
( )=
( )
( ) …
( )
(4.1)
( )
The first help function ( ) is generated by dividing the original function
(
)
with the first sub-function
. This help function will be a parabolic looking
function as shown in (4.2).
( )=
( )
(4.2)
( )
The subsequent help functions will be generated according to (4.3). The
( ) should be chosen to be feasible for hardware
sub-function
implementation.
( )
( )=
( )
(4.3)
4.1.3 Methodology for developing sub-functions
( )
The sub-functions are developed by decomposing the original function
according to second order parabolic functions in the interval 0
<1 and the subinterval within the interval. The second order parabolic function is the
decomposition function and it is easily implemented in hardware with low
complexity functions such as multiplication and addition.
The first sub-function
26
The methodology to develop the first sub-function ( ) is by dividing
( ) with as a first order approximation. There are
original function
possibilities as a result of this division, one when ( ) > 1 and one when (
(
1. The first sub-function ( ) as given by (4.4). The expression 1 +
is utilized to get approximation of these functions. The first sub-function
results in a second order parabolic function as in (4.4).
( )=
(1
1+
In equation (4.4), the coefficient
The second sub-function
)
=
+
(
)
the
two
)<
)
( )
(4.4)
is determined according to (4.5).
( )
= lim
1
(4.5)
The first help-function ( ), is important in developing the second sub-function.
The first help function ( ), is calculated according to (4.2) and dividing two
continuously concave or convex functions having the same starting and ending
point which will give rise to another function similar to a parabolic function. The
second sub-function ( ), is developed according to second order parabolic
function as shown in (4.6).
( ) =1+
(
)
(4.6)
The coefficient
in (4.6), is to satisfy the condition that the quotient between the
first help-function and second sub-function is equal to 1 when is set to 0.5 in
(4.7).
=4
1
(4.7)
The second help-function ( ) is developed in such a way that it can be divided
into a pair of functions that look like parabolic functions, where the interval of first
function is from 0
< 0.5 and for second, interval is from 0.5
< 1.
Sub-function when n>2
27
For help functions ( ) when n>2, there are one or more pairs of parabolic
looking functions. Each pair of parabolic looking function is divided into two
parabolic help-functions in order to develop higher order sub-functions. A
parabolic sub-function is developed as an approximation of the help function
( ) in the sub-interval. Sub-sub-functions are associated with a specific subinterval, and for this the subscript index is increased with the index m in sub-help( ). The numbers of sub-help-functions are doubled in every order
function in
of n>1. The corresponding sub-sub-functions are developed from these help
functions.
4.2 Hardware Implementation
An implementation for calculating real and imaginary values during the twiddle
factor multiplication in the FFT using the proposed methodology of parabolic
synthesis is described in this section.
The hardware implementation is divided into three parts; pre-processing,
processing and post-processing. Fixed point 2’s complement representation is
used.
4.2.1
Pre-Processing
Pre-processing is done to satisfy that the value of incoming angle is in the
interval 0 <1 and transfer angles from quadrant 2,3 and 4 into quadrant 1, to
achieve this operand in radians is multiplied with as shown in Fig. 4.2.
Original angle=
Normalized angle =
f
=
= (integer part)(fractional part)
(4.8)
The integer part 1 0 represents the quadrant of the original angle and it will be
used for 2’s complement conversion at the output and as MUX select bits.
Quadrant 1, 2, 3 and 4 is represented by 0, 1, 2, and 3 respectively in binary
representation as shown in Fig. 4.1.
28
Figure 4.1 Angle transformation from quadrant 2-4 to quadrant 1
The word length of the input angle is fifteen bits; one sign bit, four integer bits and
ten fractional bits to get three fractional digit precision. It is multiplied with the
factor to get the normalized angle.
Figure 4.2 Hardware architecture for pre-processing
4.2.2
Processing
The Discrete Fourier Transform (DFT) is defined as in (4.9) and (4.10)
( ) =
=
( )
(4.9)
(4.10)
The DFT is complex valued but in hardware, the real (Re) and imaginary (Im)
parts are separated in two real parts as shown in (4.11) and (4.12)
29
( )=
cos
( )=
(4.11)
sin
sin
+
(4.12)
cos
(4.11) and (4.12) can also be presented for simplicity as in (4.13) and (4.14).
=
=
× cos( )
sin( )
× sin( ) +
cos( )
(4.13)
(4.14)
The sine and cosine values must be determined by approximation. The
approximated sine and cosine functions according to first and second subfunctions (4.4) and (4.6) are in (4.15) and (4.16) respectively.
=
+
×
=1+
×
=1
+
=1+
×
(4.15)
×
(4.16)
These equations are developed according to [7] with coefficients
= 0.571
and = 0.401. Only two sub-functions are used as it will be equivalent to
precision CORDIC algorithm can deliver with 11 stages. The angle
is the
normalized fractional part of . It is only
that differs from the sin and cosine
functions. Since there is a connection between the sine and cosine terms in all 4
quadrants, the
term is used instead of
in first sub-function in the cosine
equation.
30
cos
sin
sin
2
1
2
f
f
sin
sin
2
f
2
2
cos
2
sin
1
cos
f
sin
2
sin
f
cos
2
2
1
f
sin
1
2
f
2
2
f
Figure 4.3 Input Transformation from quadrant 2-4 to quadrant 1
In figure (4.4), both sine and cosine functions are approximated and multiplied
with previous real
and
values in the FFT to get new real and imaginary
values. The normalized angle
from pre-processing block is multiplied with
is subtracted from
to get - . The
and in the next step
itself to get
and
with simple shifts and
result from - is multiplied with coefficients
add operation in Multiple Constant Multiplication (MCM) block as shown in Fig.
4.5.
- )
- ) and
There is one input - and it will generate two results
with shifts and adders. It will considerably reduce the hardware, compared to two
multipliers.
31
xim
1-
f
f
f
x
2
f
0
+
-
1
0
+
+
x
x
2's
compl.
cos
Xre
+
-
1- f +c1x( f - 2f)
2
f
+
1
0
1
x
xre
+
1+c2x( f -
2
f)
f +c1x( f -
2
f)
1
MCM unit
f
-
+
1-
f
+
1
x
0
x
+
Xim
0
1
1-
sin
2's
compl.
x
f
f
xim
Figure 4.4 Architecture for the multiplication of complex numbers with approximated
trigonometric functions (sine and cosine)
Figure 4.5 Architecture for Multiple Constant Multiplication
32
is
In case of adding “1” to
- ) term, there is no need of an adder. Since
positive, and
- ) always will be less than “1”. The fractional part can thus
be merged with “1” directly with the wiring as shown in Fig. 4.6.
Figure 4.6 The fractional bus with an added integer “1”
This wiring will give the second sub-functions for both sine and cosine functions
as shown in Fig. 4.4.
Table 4.1 shows when the input transformations are needed according to quadrant
of original angle . The integer value
is used to select the MUXes as shown in
Fig. 4.4.
Table 4.4.1 Input Transformations
1st Quadrant
Sine
Cosine
1-
f
1-
2nd Quadrant
f
f
3rd Quadrant
1-
f
1-
f
4th Quadrant
f
f
f
The upper adder with a MUX will generate first sub-function
×
for cosine function and lower adder with MUX will give
×
for sine function. The following multipliers will produce the
approximated sine and cosine values.
Since all the computation is done in the first quadrant, the output is to be
transformed back again. Table 4.2 shows the output transformations necessary at
the output.
33
Table 4.4.2 Output Transformations
1st Quadrant
2nd Quadrant
3rd Quadrant
4th Quadrant
Sine
+
+
-
-
Cosine
+
-
-
+
For 2’s complement conversion, the architecture in Fig. 4.7 is used with half
adders and XOR gates. A control signal or XOR is used to select when the
conversion is to be done according to table 4.2.
Figure 4.7 Architecture for 2’s complement conversion
After post-processing, sine and cosine, the approximated values are multiplied
with real
and imaginary
parts from previous stage of the FFT. The new
real
and imaginary
values are calculated according to (4.13) and (4.14).
34
CHAPTER
5. Results
The methodology for extracting results of area, timing and power constraints in
both the designs is shown in Fig. 5.1. The HDL description of the designs is done
in VHDL. The target design is synthesized into a gate-level netlist using an ASIC
design synthesis tool, Synopsys Design Compiler with a standard cell library such
as in our case LPLVT (Low Power Low Threshold Voltage) and LPHVT (Low
Power High Threshold Voltage). The results are taken on different voltages (1V,
1.1V, 1.2V) on each library. Area and timing information for both the designs are
gathered after the synthesis process. The synthesis process generates a Gate-level
netlist of the design and a SDC file, which contains area and timing design
constraints in the Synopsys Design Constraint format. MentorGraphics ModelSim
is used to generate toggle information of the target design in VCD (Value Change
Dump) file by using the Gate-level netlist and the SDC file from Design Compiler.
The power analysis tool, Synopsys PrimeTime estimates the static and dynamic
power dissipation in the target design using the design netlist and the VCD file.
Figure 5.1 ASIC flow for results extraction
35
5.1 Area
Area (um2)
The CORDIC and Parabolic Synthesis algorithm designs are synthesized in
Synopsys Design Compiler using STMicroelectronics 65-nm LPLVT and LPHVT
CMOS libraries. The synthesis is done using different supply voltages i.e., 1V,
1.1V and 1.2V. Fig. 5.2 and Fig. 5.3 show total area of the CORDIC and Parabolic
Synthesis algorithm designs respectively, in LPLVT and LPHVT CMOS libraries
with different supply voltages. It is obvious that parabolic synthesis design is more
area efficient than 14 stage unrolled CORDIC design.
19650
19600
19550
19500
19450
19400
19350
19300
19250
19200
19150
LPHVT
LPLVT
1
1,1
1,2
Supply Voltage (V)
Area( um2)
Figure 5.2 Total Area analysis of CORDIC design in LPLVt and LPHVt
16500
16450
16400
16350
16300
16250
16200
16150
16100
16050
16000
LPHVT
LPLVT
1
1,1
Supply Voltage (V)
1,2
Figure 5.3 Total Area analysis of Parabolic Synthesis design in LPLVt and LPHVt
36
Table 5.1 Total area of CORDIC and Parabilic Synthesis in LPLVt and LPHVt
Architecture
CORDIC
1.1V
Parabolic Synthesis
Supply Voltage (V)
1V
1.2V
1V
1.1V
1.2V
Area (um2) LPLVT
19303 19363 19370 16322 16331 16324
Area (um2) LPHVT
19502 19487 19491 16296 16304 16279
5.2 Timing
Propogation delay (ns)
Fig. 5.4 and Fig. 5.5 show the critical path propagation delay timing for the
CORDIC and Parabolic Synthesis designs respectively. It can be seen that the
propagation delay is less in parabolic synthesis compared to the 14-stage un-rolled
CORDIC design. The reason for that is the shorter critical path in the Parabolic
Synthesis design due to the high degree of parallelism. There is also a difference in
delay in two different libraries LPLVT and LPHVT. The delay is smaller in
LPLVT, that is, it is almost half of that in LPHVT.
50
45
40
35
30
25
20
15
10
5
0
LPHVT
LPLVT
1
1,1
Supply Voltage (V)
1,2
Figure 5.4 Propagation delay of CORDIC design in LPLVt and LPHVt
37
Propogation delay (ns)
35
30
25
20
15
LPHVT
10
LPLVT
5
0
1
1,1
Supply Voltage (V)
1,2
Figure 5.5 Propagation delay of Parabolic Synthesis design in LPLVt and LPHVt
Table 5.2 Total Propagation Delay of CORDIC and Parabilic Synthesis in LPLVt and LPHVt
Architecture
CORDIC
1.1V
Voltage Supply (V)
1V
Propagation delay (ns) LPLVT
23.4 19.13 16.25 13.95 11.5
9.85
Propagation delay (ns) LPHVT
47
18.66
37
1.2V
Parabolic Synthesis
26
1V
32.7
1.1V 1.2V
24.1
5.3 Static Power
Fig. 5.6 and Fig. 5.7 show the static power dissipation in the CORDIC and
Parabolic Synthesis designs respectively, using LPLVT and LPHVT CMOS
libraries. The static power dissipation is almost zero with respect to LPLVT
transistor technology. The static power dissipation in the Parabolic Synthesis is
less than in the CORDIC algorithm.
38
Leakage Power (uW)
14
12
10
8
6
LPLVT
4
LPHVT
2
0
1
1,1
1,2
Supply Voltage (V)
Figure 5.6 Static Power (Leakage) of CORDIC design in LPLVtt and LPHVt
Leakage Power (uW)
8
7
6
5
4
LPLVT
3
LPHVT
2
1
0
1
1,1
Supply Voltage (V)
1,2
Figure 5.7 Static Power (Leakage) of Parabolic Synthesis design in LPLVt and LPHVt
39
Table 5.3 Static Power (Leakage) of CORDIC and Parabolic Synthesis design in LPLVt and
LPHVt
Architecture
CORDIC
Supply Voltage (V)
Parabolic Synthesis
1
1.1
1.2
1
1.1
1.2
Static Power (uW) LPLVT
5.67
8
11.8
3
5
7.21
Static Power (uW) LPLVT
0.03 0.043 0.001 0.021 0.031 0.045
5.4 Dynamic Power
Fig. 5.8 and Fig. 5.9 describes the total dynamic power (both cell internal power
and Net Switching power) in the unrolled-CORDIC and Parabolic Synthesis
designs respectively, using LPLVT and LPHVT design libraries. The architectures
are tested on maximum frequencies.
Dynamic Power (uW)
600
500
400
300
LPHVT
200
LPLVT
100
0
1
1,1
1,2
Supply Voltage (V)
Figure 5.8 Total Dynamic Power of CORDIC design in LPLVt and LPHVt
40
Dynamic Power (uW)
450
400
350
300
250
200
150
100
50
0
LPHVT
LPLVT
1
1,1
Supply Voltage (V)
1,2
Figure 5.9 Total Dynamic Power of Parabolic Synthesis design in LPLVt and LPHVt
Table 5.4 Dynamic Power of CORDIC and Parabolic Synthesis design in LPLVt and LPHVt
Architecture
CORDIC
Parabolic Synthesis
Voltage Supply (V)
1V
1.1V 1.2V
1V
1.1V
1.2V
Dynamic Power (uW) LPLVT
267
330
404
203
250
306
Dynamic Power (uW) LPHVT
355
427
498
299
339
409
5.5 Energy
Fig. 5.8 and Fig 5.9 for dynamic power comparison gives us total dynamic power
consumed by Parabolic Synthesis and Cordic, but due to different design
approaches, both architectures have different frequencies, as shown in Fig. 5.4 and
fig. 5.5. Total energy consumed by both architectures at their maximum
frequencies is shown below in Fig. 5.10 and Fig. 5.11. It is clearly seen that
Parabolic synthesis is about 50% more energy efficient than Cordic.
41
Energy consumption (J)
1,8E-8
1,6E-8
1,4E-8
1,2E-8
1,0E-8
8,0E-9
6,0E-9
4,0E-9
2,0E-9
0,0E+0
LPHVT
LPLVT
1
1,1
1,2
Supply Voltage (V)
Figure 5.10 Energy Consumption in CORDIC design in LPLVt and LPHVt
Energy consumption (J)
1,2E-8
1,0E-8
8,0E-9
6,0E-9
LPHVT
4,0E-9
LPLVT
2,0E-9
0,0E+0
1
1,1
Supply Voltage (V)
1,2
Figure 5.11 Energy Consumption in Parabolic Synthesis design in LPLVt and LPHVt
42
Table 5.5 Energy Consumption in CORDIC and Parabolic Synthesis design
Architecture
CORDIC
Parabolic Synthesis
Voltage Supply (V)
1
1.1
1.2
1
1.1
1.2
Energy consumption (nJ) LPLVT
6.24
6.37
6.67
2.83
2.87
3.01
Energy consumption (nJ) LPHVT
16.68 15.79 12.94 9.77
8.16
7.63
43
Table 5.6.1 required memories, registers and multipliers for FFT points
No of FFT
points
No of register/Rom cells required to store
the Twiddling factors (Real + Imaginary)
Real Multipliers
16
16
12
32
32
16
64
64
20
128
128
24
256
256
28
512
512
32
1024
1024
36
2028
2048
40
44
CHAPTER
6. Conclusions
A comparison of two different algorithms, un-rolled CORDIC and Parabolic
Synthesis, has been made for area, speed and energy. These algorithms are utilized
to approximate real and imaginary values in an FFT. Comparisons have been done
on different transistor technologies at different voltages. The results show that the
parabolic synthesis algorithm is better in energy consumption, speed and area. The
area utilization in the Parabolic Synthesis is 16% less than in the un-rolled
CORDIC for the specific precision and energy consumption, on the other hand, is
50% less than the CORDIC. In the same manner, the timing of the parabolic
synthesis is about 1.4 times better than the un-rolled CORDIC algorithm. It is also
concluded that, when using a LPLVT transistor technology, the static power is
higher but less dynamic power with comparison of LPHVT transistor technology.
Since the cell size in both the technologies is almost same, the area utilization
comes to be very close to each other. Speed is better when using LPLVT transistor
technology as expected due to less parasitic effects.
An investigation of the rotations, using twiddle factor multiplications, is not
included here. However, a rough estimate on how the number of register/ROM
cells, to store the twiddle factors, increase with the number of FFT points, can be
done as shown in Table 5.1
It can be noted that the number of twiddle factor cells are increasing exponentially.
For small FFTs, registers can be beneficial but for large FFTs, A ROM will be
preferred. But the size of memory increases exponentially with increase in FFT
points, whereas on other hand the Parabolic synthesis and the CORDIC algorithms
computes rotations in the real time, hence, there is no need to store twiddle factors.
45
CHAPTER
7. Future Work
The comparison of Un-rolled CORDIC and Parabolic Synthesis algorithms for
twiddle factors can also be carried out with multiplier based FFT architecture
which needs ROM to store larger number of twiddle factors. The multiplier based
FFT architecture is area efficient for the low number of FFT points. However,
when the number of points is increasing, it utilizes an exponentially larger chip
area. Parabolic Synthesis algorithm can also be improved with respect to
wordlength optimization between the architecture.
46
References
[1] Ray Andraka, “A survey of CORDIC algorithms in FPGA based computers”,
Andraka Consulting Group, Inc. North Kingstown, USA.
[2] Y. H Hu and S. Naganathan, “An angle recording method for CORDIC
algorithm implementation”, IEEE Transaction on Computers, vol. 42, issue 1, pp.
99-102, 1993.
[3] Axler, Sheldon “Linear Algebra Done Right, 2e”, Springer Press, ISBN 978-0387-98259-5.
[4] Jack E. Volder, “The CORDIC Trigonometric Computing Technique”, IRE
transactions on Electronic Computers, vol. EC-8, issue 3, pp. 330-334, 1959.
[5] E. Hertz, P. Nilsson, “A Methodology for Parabolic Synthesis of Unary
Functions for Hardware Implementation,” Proc. of the 2nd International
Conference on Signals, Circuits and Systems, ISBN-13: 978-1-4244-2628-7, pp 16, Hammamet, Tunisia, Nov. 2008.
[6] Erik Hertz and Peter Nilsson, “Parabolic Synthesis Methodology”, In the
proceedings of the 2010 GigaHertz Symposium, pp 35, March 9-10, 2010, Lund,
Sweden.
[7] E. Hertz and P. Nilsson, “Parabolic Synthesis Methodology Implemented on
the Sine Function”, in Proceedings of the 2009 IEEE International Symposium on
Circuits and Systems, pp 253-256,Taipei, May 24-27, 2009.
47
[8] J. W. Cooley and J. W. Tukey, An algorithm for machine calculation of
complex Fourier series, Math. Comp., 19 (1965), 297-301.
[9] Kamisetty Rao, Do Nyeon Kim, Jae Jeong Hwang, Fast Fourier Transform:
Algorithms and Application, Springer Press, ISBN 976-1-40206628-3.
[10] Alan V. Oppenheim, Ronald W. Schafer, Discrete-time signal processing,
Prentice Hall, ISBN 978-0-13198842-2, the University of Michigan, Michigan,
USA.
List of Figures
Figure 2.1 A Radix-2 DIT Butterfly...................................................................... 10
Figure 2.2 A Radix-2 DIF Butterfly....................................................................... 11
Figure 2.3 A 16-point radix-2 DIF FFT algorithm.................................................. 12
Figure 3.1 Unit-length vector in Cartesian coordinate system............................. 16
Figure 3.2 Real Rotation of the vector ................................................................ 17
Figure 3.3 Real and Pseudo Rotation of the vector ............................................. 18
Figure 3.4 An un-rolled 14-stage CORDIC architecture ....................................... 21
Figure 3.5 Hardware Multiplication of scaling factor. ......................................... 22
Figure 4.1 Angle transformation from quadrant 2-4 to quadrant 1 ..................... 29
Figure 4.2 Hardware architecture for pre-processing ......................................... 29
Figure 4.3 Input Transformation from quadrant 2-4 to quadrant 1 ..................... 31
Figure 4.4 Architecture for the multiplication of complex numbers with
approximated trigonometric functions (sine and cosine).................................... 32
Figure 4.5 Architecture for Multiple Constant Multiplication.............................. 32
Figure 4.6 The fractional bus with an added integer “1” ..................................... 33
Figure 4.7 Architecture for 2’s complement conversion ..................................... 34
Figure 5.1 ASIC flow for results extraction .......................................................... 35
Figure 5.2 Total Area analysis of CORDIC design in LPLVt and LPHVt................... 36
Figure 5.3 Total Area analysis of Parabolic Synthesis design in LPLVt and LPHVt . 36
Figure 5.4 Propagation delay of CORDIC design in LPLVt and LPHVt ................... 37
Figure 5.5 Propagation delay of Parabolic Synthesis design in LPLVt and LPHVt .. 38
48
Figure 5.6 Static Power (Leakage) of CORDIC design in LPLVt and LPHVt............ 39
Figure 5.7 Static Power (Leakage) of Parabolic Synthesis design in LPLVt and
LPHVt ................................................................................................................ 39
Figure 5.8 Total Dynamic Power of CORDIC design in LPLVt and LPHVt............... 40
Figure 5.9 Total Dynamic Power of Parabolic Synthesis design in LPLVt and LPHVt
.......................................................................................................................... 41
Figure 5.10 Energy Consumption in CORDIC design in LPLVt and LPHVt .............. 42
Figure 5.11 Energy Consumption in Parabolic Synthesis design in LPLVt and LPHVt
.......................................................................................................................... 42
List of Tables
Table 3.3.1 Pre-Computed angle set .................................................................. 19
Table 3.3.2 Parameters for Generalized CORDIC ................................................ 23
Table 4.4.1 Input Transformations ..................................................................... 33
Table 4.4.2 Output Transformations .................................................................. 34
Table 5.1 Total area of CORDIC and Parabilic Synthesis in LPLVt and LPHVt ......... 37
Table 5.2 Total Propagation Delay of CORDIC and Parabilic Synthesis in LPLVt and
LPHVt ................................................................................................................. 38
Table 5.3 Static Power (Leakage) of CORDIC and Parabolic Synthesis design in
LPLVt and LPHVt ................................................................................................. 40
Table 5.4 Dynamic Power of CORDIC and Parabolic Synthesis design in LPLV t and
LPHVt ................................................................................................................. 41
Table 5.5 Energy Consumption in CORDIC and Parabolic Synthesis design.......... 43
Table 6.1.1 required memories, registers and multipliers for FFT points............. 44
49
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
advertisement
© 2021