Sine Function Approximation using Parabolic Synthesis and Linear Interpolation

Sine Function Approximation using Parabolic Synthesis and Linear Interpolation
Master’s Thesis
Sine Function Approximation using
Parabolic Synthesis and Linear
Interpolation
By
Madhubabu Nimmagadda
Surendra Reddy Utukuru
Department of Electrical and Information Technology
Faculty of Engineering, LTH, Lund University
SE-221 00 Lund, Sweden
Abstract
This Thesis work deals with the ASIC implementation of the sine function
approximation using the combination of Parabolic Synthesis methodology and
linear interpolation. Parabolic Synthesis is a novel methodology for
implementing unary functions such as logarithmic, trigonometric, and
arithmetic functions. This methodology gives a high degree of parallelism and
efficient use of hardware compared to existing algorithms like polynomial
approximation, CORDIC etc. Parabolic Synthesis is combined with linear
interpolation in order to achieve the high accuracy. The architecture is
designed to achieve increased performance, lesser area and lower power
consumption.
Keywords: Parabolic Synthesis, Linear interpolation, MCM, Sine function
2
Acknowledgments
It is a pleasure to thank everyone who has made this dissertation possible and
helped us throughout the project with their excellent guidance. We would not
have finished our thesis without their support and motivation.
First, we would like to express our gratitude to Professor Peter Nilsson, who
has provided the thesis and let us experience in the field of Digital ASIC. We
have had good guidance, patience, caring, and very pleasant work atmosphere.
Second, our deepest thanks to our supervisor Erik Hertz, who had given the
best suggestions and solutions to get out of the problems faced during our
entire project. He had given us his valuable assistance and appreciation toward
the work we have done. We will never forget his smile and his friendly nature
toward explaining, making doubts clear. We are thankful to the Department of
EIT, who has provided and given the access to the most expensive project
tools.
Finally, we would like to thank our friends and parents, who always support us
in our life and keep up our happiness.
Madhubabu
NimmagaddaSurendra Reddy
Utukuru
3
Table of Contents
Abstract............................................................................................................................2 Acknowledgments........................................................................................................3 TableofContents..........................................................................................................4 Preface..............................................................................................................................6 1 Introduction..........................................................................................................7 2 ParabolicSynthesis............................................................................................9 2.1 Introduction................................................................................................9 2.2 Normalization.............................................................................................9 2.3 Methodologydevelopment..................................................................10 3 Interpolation.......................................................................................................13 3.1 Introduction..............................................................................................13 3.2 Interpolationmethods...........................................................................13 3.3 Linearinterpolation...............................................................................14 4 Architecturedevelopment.............................................................................15 4.1 Introduction..............................................................................................15 4.2 Analysis.......................................................................................................15 4.2.1 Onesub‐functioncombinedwithlinearinterpolation...................16 4.2.2 Twosub‐functioncombinedwithlinearinterpolation..................17 4.2.3 Comments.........................................................................................................18 4.3 Developmentofarchitecture..............................................................18 4.3.1 ParabolicSynthesisarchitecture.............................................................18 4.3.2 Linearinterpolationarchitecture...........................................................21 4.3.3 Fourquadrants...............................................................................................22 5 Hardwareimplementationandoptimization.........................................24 5.1 Designflow.................................................................................................24 5.2 Toplevelarchitecture............................................................................25 5.3 Parabolicsynthesisarchitecture.......................................................26 5.3.1 Architecture.....................................................................................................26 5.3.2 Processing.........................................................................................................26 5.3.3 Hardwareoptimizations.............................................................................27 5.4 Linearinterpolationarchitecture.....................................................28 5.4.1 Overview............................................................................................................28 5.4.2 Hardwareoptimizations.............................................................................28 5.5 Post‐processing........................................................................................29 5.6 Wordlengthevaluation.........................................................................30 6 Results...................................................................................................................35 6.1 Testsetup...................................................................................................35 4
6.2 Synthesisresults......................................................................................36 6.2.1 Designconstraints.........................................................................................36 6.2.2 Areaofdifferentmodulesinthedesign...............................................38 6.2.3 Maximumclockfrequency.........................................................................38 6.3 TimingandAreaanalysis.....................................................................39 6.3.1 Timing.................................................................................................................39 6.3.2 Area......................................................................................................................39 6.4 Poweranalysis..........................................................................................40 6.4.1 DynamicPower...............................................................................................40 6.4.2 StaticPower.....................................................................................................42 6.5 Energy..........................................................................................................44 6.6 Comparison................................................................................................45 7 Conclusion...........................................................................................................47 8 Futurework........................................................................................................48 9 References...........................................................................................................49 10 ListofFigures.................................................................................................51 11 ListofAcronyms...........................................................................................52 5
Preface
In this thesis work, we both have worked together except from the literature work,
and total work has been scheduled based on our availability. We have divided the
tasks and worked together throughout the project.
This thesis documentation is organized into 8 chapters. Chapter 1 presents a short
introduction to exist algorithms along with their merits and demerits. Chapter 2
provides description of Parabolic Synthesis methodology and its advantages over
other algorithms. Chapter 3 explains the linear interpolation technique and chapter
4 contains a detailed description of the architecture development procedure for the
parabolic synthesis methodology and linear interpolation approximation.
Architecture implementation of design in hardware is explained in chapter 5. In
chapter 6, Design flow and results from the hardware implementation are
discussed and compared with other implementations. Conclusion and future work
is explained in chapter 7 and 8.
6
CHAPTER
1
1 Introduction
Unary functions like Logarithmic, Exponential, and Trigonometric functions as
well as Arithmetic functions [1] plays an important role in applications like
wireless systems, computer graphics, digital signal processing. For the
implementation of these functions we have to choose an algorithm, which is good
and plays an important role in area, computational speed, and accuracy. The
hardware implementation of such functions is a challenge to meet the desired
requirements.
There exist different algorithms such as single lookup table, polynomial
approximation, and CORDIC algorithm, which has own merits and demerits. A
single lookup table [2] is an easy and direct method in implementation. However
this method suits for the low precision applications, not for low area applications.
A Single lookup table can be approximated by using interpolation technique in
order to increase the accuracy up to an extent.
Polynomial approximations [3] are performed with additions and multiplications.
The computational complexity will be depending on which polynomial schemes
you will choose in the existed schemes. Depending on the requirements of the
design, polynomial approximation can be implemented either with least squares
approximation or with least maximum approximation.
Another method is the CORDIC algorithm [4], which is good in efficient use of
hardware since it has no multipliers. The CORDIC algorithm is an iterative
algorithm, which is implemented by using shifts and additions. Increased number
of iterations increases the bit accuracy but makes the algorithm unfeasible for
high speed applications.
In order to overcome some limitations in the previous methods, a new algorithm
called Parabolic Synthesis [5] has been chosen. It is a novel methodology that
came up with a very fast computational techniques and better accuracy, when
implementing in hardware. In this Master thesis work, the design of the sine
function approximation is done by using Parabolic Synthesis methodology and
7
followed by linear interpolation approximation aiming for better accuracy, area,
and power consumption.
8
CHAPTER
2
2 Parabolic Synthesis
2.1 Introduction
Parabolic Synthesis is a different methodology to implement functions like
Trigonometric, Logarithmic, and Arithmetic functions, which plays a vital role in
the field of signal processing, image processing, and robotics, where the speed of
computations has an important role [5].
The Parabolic Synthesis methodology makes use of an original function to develop
the sub functions and help functions, which are implemented in the hardware [5].
The proposed architecture possesses a high level of parallelism and consumes less
time in computation. The design is built only with multipliers, adders and shifters,
which is easily implemented in hardware.
The Parabolic Synthesis methodology works better in getting efficient response in
terms of area, power and time. The main steps in this methodology are
normalization, approximation and transformation [5].
2.2 Normalization
Normalization is the first step in Parabolic Synthesis methodology, which limits
the input in the numerical range 0 to 1.The inputs and output for a particular unary
function ,which is notated as forg(x) ,should be limited to 0 to 1 on both x and y
axis. The function should be either convex or concave throughout the interval.
9
2.3 Methodology development
The Parabolic Synthesis methodology is developed in approximation of any unary
function, it always good to think in hardware perspective so that they can be
synthesized with good metrics. The simple operations like multiplication, addition,
and shifts are chosen in designing architecture.
,
The proposed methodology will be derived in to a set of sub functions
which are recombined in order to get an approximated original function.
∗
….
.
represents a parabolic looking function, which is called first help
In (2.1),
function. The use of number of sub functions closes the approximated value to
original value and hence better accuracy. The sub functions derived as parabolic
looking functions should be feasible, so that they can be implemented in hardware.
The second sub function is developed by using first help function
.
2.2
In the same manner, the other help functions are also developed in order to
develop other sub functions as shown in (2.3).
2.3
Developing sub-functions
in Parabolic Synthesis methodology is developed by
The first sub function
, with x as first order approximation. The two
dividing the original function
possibilities after the division is either
> 1 or
< 1.To develop the first sub
function,1
1
is used an approximated term. The multiplication of the
approximated term with x gives the first sub-function [1] [16].
1
1
∗
2.4
The coefficientc1 in (2.4) is calculated determined according to (2.5)
lim
→
1 2.5
10
The first help function is calculated according to (2.2) and is used in developing
the second sub function, which is also a parabolic function.
1
2.6
in (2.6) is chosen to satisfy that the quotient between the first help function
fhelp1(x) and the second sub-function s2(x) is equal to 1 when x is set to 0.5.
4
1
2
1 2.7
The sub functions n> 2 are developed by dividing the help function into sub
intervals as equated in (2.8). The parts of help functions are used to develop
further partial sub functions [1].
,0
,
1
2
1
2
,
2
2
………………….………
2
1
,
,
2
2.8
,
1
is used instead of x in partial sub functions in
The normalized value of input
order to have a feasibility to implement them in hardware. The partial sub
functions are notated as (2.9) [1] [16].
,0
,
1
1
2
2
,
2
2
………………….………
2
1
,
,
2
2.9
,
1
The recombination of partial sub functions in (2.9) gives an approximation of
and parabolic sub function is formulated as (2.10)[1] [16].
,
1
,
∗
(2.10)
Coefficients in parabolic sub functions are calculated according to (2.11)
11
,
4∗
,
∗
2∗
1
1
2
1 2.11
In parabolic synthesis methodology, to implement any unary function using more
than two sub functions results more hardware. The third sub function uses two
coefficients and needs one more multiplier to recombine, which is a demerit in
hardware perception. It is always a tradeoff to choose how many bits of accuracy
are desired and how many number of sub functions should be used.
12
CHAPTER
3
3 Interpolation
3.1 Introduction
The interpolation concept is used to approximate the function with the known data
points by analyzing the numerical data [9]. There are several ways of interpolating
a continuous signal like Linear, Piecewise, and spline. From all the existing
interpolation techniques, linear interpolation is having its own advantages which
possess speed, simplicity, and accuracy [9].
3.2 Interpolation methods
There are several interpolation methods existed and every method is having its
own merits and demerits. Depends on application and implementation, user has to
make a tradeoff to choose which method is best suited. Piecewise constant
interpolation, Linear interpolation, Spline, and the Polynomial interpolation are
different interpolation methods, which are commonly used in approximation of
functions.
Piecewise constant interpolation or nearest neighbor interpolation is the easiest
and the simplest method among all the existing methods [10]. In this technique,
the method just locates the nearest data value and assigns to it. This method is
chosen for image processing applications where multivariate interpolation is
needed [9].
Linear interpolation is the other method, which uses some known data points to
calculate an unknown data point [9]. This method is well suited for the
applications like computer graphics and easily implemented with Lookup tables.
To work in other spatial domains, one can go for bilinear interpolation or trilinear
interpolation methods.
Polynomial interpolation is another form of interpolation, where the interpolant is
a polynomial function, which is a linear equation in case of linear interpolation
[11]. Here the Interpolation error is proportional to the distance between the data
points and interpolant is indefinitely differentiable, since it is a polynomial [11].
On the other hand, the polynomial interpolation has a drawback of computational
complexity and oscillatory artifacts [11]. The spline interpolation method
13
overcomes the disadvantages of polynomial method by using a low degree of
polynomial. The computational complexity is also lower with same accuracy as
the polynomial interpolation [11].
3.3 Linear interpolation
Linear interpolation is a method of finding new data points (x, y) from the known
data points (x0 y0, x1y1) as shown in Fig. 1. The new set of data points (x, y) can be
found by the slope of straight line between two known points [9].
Figure 1. The Method of interpolation
As shown in Fig. 1, the value of y for a known value of x is calculated according
to (3.1).
∗
3.1 Elaborating (3.1), y0 is the value from where the interpolation starts and the second
part in the product term is the gradient value of the line. Linear interpolation with
different intervals like 8, 16, 32 and 64 are chosen to get desired accuracy and
precision of values.
14
CHAPTER
4
4 Architecture development
4.1 Introduction
In this section, architecture development for sine function using Parabolic
Synthesis methodology combined with linear interpolation is explained. As
explained in section 2, Parabolic Synthesis methodology provides feasibility in
implementation of unary functions in accuracy and hardware utilization also.
Providing some optimizations for architecture also makes more efficient hardware
implementation. In this case, sine function is the original function forg (x) and is
used to develop the sub functions. There are several algorithms to implement a
sine wave in hardware as explained in section 1, every algorithm is having its own
merits and demerits. In this project, a novel methodology Parabolic Synthesis is
chosen to gain the less computation complexity and speed as well. Parabolic
Synthesis is a methodology, which is constructed with different stages called as
sub functions to compute value for any unary function [5]. The Error from these
sub functions is analyzed and approximated by using linear interpolation
technique.
An analysis is made to choose the number of sub functions to be used in
achieving more than 15 bits of accuracy. The third sub function is implemented
with linear interpolation technique, which gives a better result in all the metrics.
The objective of the design is achieving better accuracy with certain input
constraints, less power consumption and less computational speed.
4.2 Analysis
The Parabolic synthesis architecture consists of different sub functions and the
product of them gives an approximated original function. The accuracy of the
output increases as the number of sub functions is increased. Increasing the
number of sub functions consumes more hardware area and computational speed.
So, an analysis is made for architecture to use how many sub functions are to be
combined with linear interpolation in order to achieve the accuracy desired. The
accuracy is calculated by measuring the difference between original function and
derived function.
20 ∗ log
15
(4.1)
In dB scale 6.02 dB represents 1 bit accuracy. For example, 61 dB represents 10
bit accuracy. In this case, two different scenarios are considered to analyze the
absolute error and accuracy, which are discussed further.
4.2.1 One sub-function combined with linear interpolation
The sine function approximation is designed with one sub function combined with
different levels of interpolation to analyze the bit accuracy. The design is
approximated in the range 00<x<900and 8, 16, 32, and 64 levels of interpolation is
used.
Figure 2. The Error analysis for 1 sub function & linear interpolation
In Fig. 2, the error analysis is shown for different scenario in dB value. The
resolution in bits is as below.
16




8 intervals, resolution 9.1 bits
16 intervals, resolution 11.1 bits
32 intervals, resolution 13.09 bits
64 intervals, resolution 15.1 bits
As the number of interpolation levels is doubled, the resolution is increased by 2
bits. In order to achieve 15 bit accuracy, a 64 level interpolation is needed and in
order to implement in hardware, a Look up Table (LUT), which can store 64
different values should be used.
4.2.2 Two sub-function combined with linear interpolation
As explained in 1.2.1, in order to achieve more than 15 bit accuracy, it requires
more hardware area with one sub function combined with linear interpolation. In
this section, the sine function is approximated by using the 2 sub functions
combined with linear interpolation. The design is approximated in the range
00<x<900and 8, 16, 32, and 64 levels of interpolation is used.
Figure 3. The Error analysis for 2 sub functions & linear interpolation
In Fig. 3, the error analysis for sine function approximation with 2 stage parabolic
synthesis combined with linear interpolation is shown in dB scale. The resolution
for different levels of interpolation is as below.


8 intervals, resolution 11.7 bits
16 intervals, resolution 13.4 bits
17


32 intervals, resolution 15.6 bits
64 intervals, resolution 17.5 bits
The sine function approximation designed with 2 stage parabolic synthesis
combined with 32 level interpolation gives more than 15 bits of accuracy.
4.2.3 Comments
The approximation of the sine function implemented with parabolic synthesis
methodology provides feasibility to implement in hardware. The linear
interpolation approximation technique when combined to Parabolic Synthesis
provides good results in accuracy, chip area, and computational speed as well. As
the number of levels of interpolation is doubled, the resolution and error
distribution is increased by 2 bits. Changing the coefficients, in linear interpolation
manually, can increase the bit accuracy. The implementation of linear
interpolation methodology for the approximation of sine function requires
memories and more access time. Therefore, the concept of a 2-stage Parabolic
Synthesis combined with linear interpolation is the feasible way to achieve less
size memories and less access time.
4.3 Development of architecture
4.3.1 Parabolic Synthesis architecture
The Parabolic Synthesis methodology makes use of an original function to develop
the sub functions and help functions, which are implemented in the hardware [5].
The proposed architecture possesses a high level of parallelism and consumes less
time in computation. The design is built only with multipliers, adders and shifters,
which is easily implemented in hardware.
The Parabolic Synthesis methodology works better in getting efficient response in
terms of area, power and time. The main steps in this methodology are
normalization, approximation and transformation [5].
4.3.1.1 Normalization
The Parabolic Synthesis methodology for a specific unary function can be
computed mainly in three steps, normalization, approximation, and transforming.
In Normalization, the numerical range will be limited to a range in order to
simplify the hardware implementation. For any Unary function y=forg (x), both x
axis and y axis should be in range of 0 and 1. Fig. 4 shows sine function after
normalization.
18
Figure 4. The Normalization for Sine function
4.3.1.2 Calculation of coefficient in sub-functions
The hardware is completely designed with low complexity operations like
multiplications, additions and shifts when approximating the original function. The
proposed methodology uses the basic function to formulate the sub functions and
later multiplication of the sub functions gives the approximated original function.
The methodology is constructed with second-order parabolic functions [5] named
as sub functions and recombination of these sub functions gives approximation of
original function forg (x), in this case the sine function, as shown in (2.1). The
accuracy will be better, if more sub functions are used in the calculation.
The help function generated at each stage is used in calculating the next sub
function. For example as shown in (2.2), the division of the original function by
the first sub function gives the first help function, which is used when deriving the
second sub function. The subsequent help functions are generated from the
according to (2.3).
When developing the sub functions, the original function is decomposed into
second-order parabolic functions for the interval 0 to 1. The second order
parabolic functions are chosen to avail a benefit of implementing with less
complexity in hardware. The third sub function from the Parabolic Synthesis
methodology contains two coefficients which increases the hardware complexity
[9]. The accuracy will be better by replacing the third sub function with another
approximation technique.
19
The first sub function is calculated by dividing the original function with x as a
first order approximation, which is specified in (2.2). The result have two
possibilities, fdiff(x)>1 and fdiff(x) <1 [14]. In order to approximate the original
function, f2(x) is developed from fdiff (x), which is specified in (2.3). The first sub
function is calculated by multiplying x and f2(x) as shown in (4.3). The related plot
in developing first sub function is shown in Fig.5.
lim
1
→
1
1
0.5703125 4.2
4.3
where c1 is specified in (4.2)
∗
4.4
Figure 5. The Developing the first sub function
The first help function, calculated from (4.5) is used when deriving the second sub
function. The second sub function s2(x) is calculated from (2.6).
4.5
c2 in (2.6) is chosen to satisfy that the quotient between the first help function
fhelp1(x )and the second sub-function s2(x) is equal to 1 when x is set to 0.5.
20
4
1
2
1 0.400390625 4.6
Figure 6. fhelp (x) & second sub function s2(x)
Fig. 6 shows plots for the help function and the second sub function. The crossing
at x=0.5 can be clearly seen from the graph.
4.3.2 Linear interpolation architecture
In case of the sine wave approximation, linear interpolation is applied by
calculating the gradient values at different levels. Equation (4.7) gives the third
sub function, which is used for the linear interpolation technique.
∗
4.7 From the equation, K is the value from where the interpolation is starting and G is
the slope of the function. The most significant bits from the input x are used to
index the particular gradient value and K value.
The values of K and G are measured from the help function generated by dividing
the original function and result from Parabolic Synthesis as shown in (4.8).
4.8
21
The value of K is derived from the help function and depends on the number of
levels in the interpolation. For n levels of interpolation, n number of K values is
chosen from the help function at successive time intervals. The gradient (G) value
is calculated by subtracting the successive K values according to (4.9).
1
4.9
Fig. 7 shows the plot of the K values measured from the help function.
Figure 7. K values for the third sub function
4.3.3 Four quadrants
The calculations are developed for the first quadrant and extended to calculate the
sine values in other three quadrants also. The original angle is normalized and
transformed into first quadrant by using a factor 2/ π.
∗
4.10)
2/π
ᶱ
X
ᶱf(integer part , fraction part )
Figure 8.Normalization of original angle
22
For example, 2π/3 will be represented as
2
4
1
2
01 4.11 3
3
3
The result in (4.11) is divided into integer part and fractional part. The integer part
and is used by the multiplexer in the
is represented using two binary bits
architecture. These two select lines represent the quadrant of the original angle.
The same bits are also used in two’s complement representation of the output
when sine is in 3rd and 4th quadrant. The changes in architecture in view of the
hardware are explained briefly in next chapter. The architecture for all four
quadrants is possible by inserting one multiplexer and a two’s complement
converter.
Table 1. Input transformation for 4 quadrants
Quadrant
1
Theta_sel(
00
2
01
3
10
4
11
The table 1 represents the
are selected by multiplexer.
)
Input value
1
1
values for all the quadrants and input vales, which
23
CHAPTER
5
5 Hardware implementation and
optimization
5.1 Design flow
The implementation of the design in hardware starts with making a reference
model in MATLAB and verifying the functional behavior of the design. This
thesis is organized similar to the ASIC flow as shown in Fig. 9. After a
successfully working reference model, the design is coded in VHDL and simulated
with the Modelsim tool. Furthermore, the results are compared with reference
model. The design is synthesized for generating a netlist and is used to analyze the
power consumption. All possible switching activities are generated as Value
Change Dump (VCD) files using Modelsim, which is used by the Primetime tool
for power analysis.
Figure 9. The Design flow
24
5.2 Top level architecture
In this chapter, the global architecture for complete design is explained in top
down methodology. Two stage parabolic synthesis, linear interpolation,
optimization techniques, and word length evaluation are explained briefly.
Fig. 10 shows the top level architecture for the design. The architecture is divided
in to two parts, which is represented in dashed lines. The upper part represents a
two-stage Parabolic Synthesis and lower part represents a linear interpolation.
Several optimization techniques are applied to improve the hardware efficiency.
The complete architecture is designed and implemented with 4 multipliers, 7
adders, 1 multiplexer, and a 2’s complement module. The detailed description of
every stage is explained further.
Figure 10. The Top level architecture
25
5.3 Parabolic synthesis architecture
5.3.1 Architecture
The implementation of sub functions in Parabolic Synthesis and optimization
techniques in the architecture is discussed in this section. The implementation is
divided into 3 steps namely preprocessing, processing, and post processing, which
are discussed further briefly. Fig. 11 represents the architecture for parabolic
synthesis.
Figure 11. The Architecture of 2 Stage Parabolic Synthesis
The design consists of a multiplexer to select the inputs for a specific quadrant,
and Multiple Constant Multiplier (MCM) unit to replace the multipliers with
adders. This architecture results the product of two sub functions, which is
multiplied to the result from the linear interpolation to get the final output.
5.3.2 Processing
Processing is the main part of the architecture, which calculates the product of two
sub functions. The two parabolic functions in the methodology are easily
implemented in hardware since they only need normal adders and multipliers.
Applying some optimization techniques efficiently can reduce the number of cells
required. Equations (2.4) and (2.6) are the two parabolic functions to be
implemented in hardware and product of those two sub functions results in the
output.
∗
5.1
The front end part of the design is a multiplexer, which selects the input depends
on selection lines. The constants
and
calculated in (4.2) and (4.6)
respectively are used and it’s equivalent binary representation is shown below.
26
0.5703125
0. 0.400390625
10010010
5.2
01100110
5.3
5.3.3 Hardware optimizations
Multipliers plays an important role in occupying the hardware area. Different
optimization techniques have to be applied to avoid the usage of multipliers. In the
design, two multipliers have the common multiplicand with
and
as
multipliers. The MCM unit replaces those two multipliers with 4 adders. The
MCM unit is a replacement for multipliers when they have common multiplicand,
which is built with shift operations and adders, which results less hardware area.
Fig.12 shows the implementation of MCM unit with 4 adders. The MCM unit
takes the value
and generates the product value with both c1 and c2.
IN
>>6
>>3
C1(x‐x2)
>>2
>>7
C2(x‐x2)
Figure 12. The Multiple Constant Multiplier (MCM) unit
27
The other optimization in hardware is the replacement of an adder in calculating
the second sub function. In the calculation of the second sub function, the value 1
is added to c2* (x-x2). The adder in this can be removed and a binary ‘1’ is padded
at MSB position as presented in Fig. 11.
5.4 Linear interpolation architecture
5.4.1 Overview
In this chapter, the architecture for the third sub function with linear interpolation
is explained. Fig. 13 shows the architecture for the third sub function that consists
of
1 multiplier, 1 adder, and 2 Look Up Tables (LUT).
Figure 13. The architecture of the third sub function (Linear interpolation)
In the design, the values K and G are calculated from (4.3).They are stored in a
LUT and the MSB of the input are used to address the position of the bits.
5.4.2 Hardware optimizations
The values of G from table 2 contain both positive and negative values. If they
have to be used in 2’s complement representation with same precision requires
more number of 1’s which results to more power consumption and chip area as
28
well. Providing some logic in the arithmetic calculations avoids this kind of
representation. In this case, a dual functional adder is built, which can perform
both addition and subtraction.
K
G*Xlsb
Addsub
K +/‐ G*Xlsb
Sign(1 bit)
Figure 14. The Addsub Component
As shown in Fig. 14, sign bit decides the operation should be performed by the
component. The first MSB in G is used to represent either if the value is positive
or negative.
Hardware implementation of third sub function needs a multiplier and an adder.
Fig.14 shows the hardware architecture of the linear interpolation method. The
adder in this case is a dual functional, which means it can be switched to have
either an addition or subtraction functionality. We needed this since, the product of
G and
can be both negative and positive. Fig.14 shows the Addsub
component. Because of sine function’s symmetry property and by reading in
reverse order, the K and G coefficients stored in LUTs can be used for all the four
quadrants.
The output from linear interpolation is multiplied with the parabolic synthesis
output to get the final result.
5.5 Post-processing
The value of sine function is negative in the
en third and fourth quadrants. The post
processing part transforms the output in-to a feasible form. The results in negative
are converted into 2’s complement representation. The 2’s complement component
is switched only in 3 and 4 quadrants and q1 is used to enable or disable it. Fig. 15
shows block diagram of 2’s complement converter.
Input
2’s complement
Figure 15. Two's complement module
29
2’s complement
output
5.6 Word length optimization
The design is aimed for the usage of fewer bits to represent the data at different
stages. The study to fix the word length at every stage is done in MATLAB and is
verified, as well as for the loss of precision. Fig. 16 shows the block diagram with
word length at different stages.
Figure 16. Block diagram with wordlengths
30
Table 2. Word lengths at every stage
Name
Square
x_sub
MCM
Input1(length) Input2(length)
xi(14)
xi(14)
xi- xi2(15)
xi(14)
xi2(15)
---------
Original
output(length)
xi2(28)
xi- xi2(15)
c1(7) * xi- xi2
c2(7) * xi- xi2
s1 _ add
xi(14)
c1 * (xi- xi2 )
(17)
s2 _ add
1
c2 *( xi- xi2 )
(17)
s1 s2 mult
s1(18)
s2(18)
G_X_Mult
K
_GX_Add
S3_ fp
_Mult
xLSB(9)
K (19)
G (19)
G * xLSB (19)
S3 (19)
fp (19)
s1 = xi +c1 *
(xi- xi2 )
(18)
s2 = 1 +c2 *
(xi- xi2 )
(18)
fp = s1 * s2
(36)
G * xLSB(28)
S3=K + G *
xLSB (19)
S3 * fp (38)
Truncated(lengt
h)
xi2(15)
xi- xi2(15)
c1 * (xi- xi2 )
(17)
c2 *( xi- xi2 )
(17)
s1 = xi +c1 * (xixi2)
(18)
s2 = 1 +c2 * (xixi2 )
(18)
fp = s1 * s2
(19)
G * xLSB(19)
S3 = K + G *
xLSB (19)
S3 * fp (19)
The input which is in the range 0 to 1 is represented using 14 plus 2 bits used for
selecting the transformation of the input and output in case of the other quadrants.
A trial and error method to fix the word lengths at every stage has been used. A
function is developed in MATLAB to truncate the word length at each stage in a
flexible way. A generic methodology is chosen in designing the components even
in implementing in the whole design in hardware.
In Fig. 16, the word lengths at every stage are shown. The data will be truncated to
a specific word length after every multiplier, since the multiplier increases the
number of bits. The accuracy desired is cross verified when truncating the bits at
every stage. The MUX in architecture selects any one of the 14-bit inputs using
two selection lines. The selection of the inputs is needed to calculate the sine value
in other quadrants. The squarer unit squares the input and the word length of the
input is doubled to 28 bits. The xi2 is truncated to 15 bits from 28 bits. The xi- xi2 is
calculated using a subtractor. No data truncation is needed after any adders or
subtractor since they does not increase word length to a bigger number , but before
the operation the word lengths should be balanced in magnitude by adding zeros at
LSB position, which is represented as ‘&0’ in Fig. 16 . The MCM unit is designed
to calculate the two product terms specified in table.4 according to the coefficients
c1 and c2 , which are 7 bit length. To calculate the first sub-function s1 ,the input xi
31
is added to one of the product from MCM unit. To calculate the second subfunction s2 , to add a value of 1,one more bit is padded to the other product term at
MSB position as ‘1&’ denoted in Fig. 16.
The multiplier S1S2_mult gives the product of two sub functions in 35 bit size,
which is then truncated to 19 bits. The LUT contains K and G values, which are 19
bit word length, which are tabulated in table 3 and table 4. The LSB 9 bit in x is
multiplied with G and the result is 28 bits in length, which is then truncated to 19
bits. K_gxadd, one more adder adds the K value to the product and results third
sub function with linear interpolation approximation. The product of two sub
functions and the third sub function are multiplied using a multiplier named as
S3_fp_pro to produce the final result. The result is truncated to 19 bits from 38
bits. The truncation at each stage is a trial and error method keeping in mind that
the desired accuracy is more than 15 bits with minimum word lengths. The word
lengths of inputs and outputs at every stage and every component are tabulated in
table 2.
32
Table 3. Values of K in Linear interpolation
Sl. no
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
Floating value
1
0.999275207519531
0.998470306396484
0.997886657714844
0.997474670410156
0.997238159179688
0.997127532958984
0.997154235839844
0.997272491455078
0.997489929199219
0.997768402099609
0.998119354248047
0.998504638671875
0.998935699462891
0.999374389648438
0.999839782714844
1.000293731689453
1.000747680664063
1.001171112060547
1.001579284667969
1.001934051513672
1.002258300781250
1.002513885498047
1.002716064453125
1.002834320068359
1.002880096435547
1.002822875976563
1.002677917480469
1.002407073974609
1.002025604248047
1.001499176025391
1.000835418701172
Binary equivalent
1000000000000000000
0111111111101000010
0111111111001101111
0111111110111010110
0111111110101101010
0111111110100101100
0111111110100001111
0111111110100010110
0111111110100110101
0111111110101101110
0111111110110110111
0111111111000010011
0111111111001111000
0111111111011101001
0111111111101011100
0111111111111010110
1000000000001001101
1000000000011000100
1000000000100110011
1000000000110011110
1000000000111111011
1000000001001010000
1000000001010010011
1000000001011001000
1000000001011100111
1000000001011110011
1000000001011100100
1000000001010111110
1000000001001110111
1000000001000010011
1000000000110001001
1000000000011011011
33
Table 4. Values of G in Linear interpolation
Sl. no
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
Floating value
-7.209777832031250e-04
-8.049011230468750e-04
-5.836486816406250e-04
-4.100799560546875e-04
-2.384185791015625e-04
-1.068115234375000e-04
2.288818359375000e-05
1.182556152343750e-04
2.174377441406250e-04
2.784729003906250e-04
3.490447998046875e-04
3.852844238281250e-04
4.291534423828125e-04
4.405975341796875e-04
4.615783691406250e-04
4.520416259765625e-04
4.539489746093750e-04
4.234313964843750e-04
4.062652587890625e-04
3.566741943359375e-04
3.204345703125000e-04
2.555847167968750e-04
2.021789550781250e-04
1.163482666015625e-04
4.577636718750000e-05
-5.531311035156250e-05
-1.468658447265625e-04
-2.689361572265625e-04
-3.814697265625000e-04
-5.264282226562500e-04
-6.618499755859375e-04
-8.354187011718750e-04
34
Binary equivalent
(1) 101111010
(1) 110100110
(1) 100110010
(1)0 11010111
(1)00 1111101
(1)0000111000
(0)0000001100
(0)0000111110
(0)0001110010
(0)0010010010
(0)0010110111
(0)0011001010
(0)0011100001
(0)0011100111
(0)0011110010
(0)0011101101
(0)0011101110
(0)0011011110
(0)0011010101
(0)0010111011
(0)0010101000
(0)0010000110
(0)0001101010
(0)0000111101
(0)0000011000
(1)0000011101
(1)0001001101
(1)0010001101
(1)0011001000
(1)0100010100
(1)0101011011
(1)0110110110
CHAPTER
6
6 Results
6.1 Test setup
The design is verified with the help of a test bench by stimulating the inputs
from a text file. The output from the top module is written to a text file,
which is imported to MATLAB to compare it with the reference output.
Fig. 17 shows the entire setup for testing the design.
CLK
Input
RST
Thetasel[1:0]
TestBench
Input[13:0]
Top
Design
Output
Output[18:0]
MATLAB
Verification
Figure 17. The Design Test setup
The absolute error for the original function and results from the hardware
implementation are calculated to check the accuracy. Fig.18 shows the absolute
error calculated for 1 sub function, 2 sub functions, and 3 sub functions. The
accuracy is increased to more than 15 bits as the number of sub functions are
increased to 3.
35
Figure 18. The absolute error of the original function compared with the result using
1, 2, and 3 sub functions
6.2 Synthesis results
Synthesis is one of the important and the basic step in the ASIC flow. Here the
design tool generates gate level net lists from the Register Transfer Level (RTL).
A design tool called design vision or DC-compiler by Synopsys has been used for
the synthesis. The netlist is generated by using STMicroelectronics 65nm CMOS
technology. Different parameters of the design such as area, time and maximum
clock frequency are analyzed. Further, this generated netlist is used for power
analysis and will be explained in the chapter 6.3.
6.2.1 Design constraints
Low Power Low Threshold Voltage (LPLVT) and Low Power High Threshold
Voltage (LPHVT) are the two different technology libraries that are used to
synthesize the design. The design is synthesized using two specified libraries for
different supply voltages to estimate both static and dynamic behavior. Maximum
speed and minimum area are the two design constraints to the design. One is to
find critical path in the design and the other to know minimum area. These
constraints are applied to the design during synthesis.
6.2.1.1 Maximum Speed
The maximum speed of the design estimated by setting up the clock to 1ns, area
has not set to any value, LPHVT libraries and 1.2V as supply voltage. As seen
from the table 5, we have got negative slack for maximum speed optimization. A
slack of 4.29ns is added to the clock to know the critical path time.
36
Table 5 .Maximum speed optimization
MEASURES
Data Arrival Time
Data Required Time
Critical path slack
Combinational Area
Non-combinational Area
PARABOLIC
SYNTHESIS
5.16
0.87
-4.29
19000
367
Total Area
19400
UNITS
ns
ns
ns
6.2.1.2 Minimum Area
Minimum area of the design is estimated by setting up the set_maxarea to zero
during synthesis. The design is synthesized for LPHVT libraries and 1.2V as
supply voltage. From table 6, we can see that the tool optimizes more on area
compared to maximum speed.
Table 6 .Minimum area optimization
MEASURES
PARABOLIC
SYNTHESIS
UNITS
Data Arrival Time
Data Required Time
Critical path slack
Combinational Area
Non combinational Area
19.83
-17.93
1.86
11000
343
ns
ns
ns
Total Area
11400
Area and speed are the two major features of the design to be interested. The table
5 and 6 are the two different synthesis runs that are necessary for an area
optimized design and speed optimized design. The set_max_area to zero applies
area constraint to the design from which the area cost of the design is calculated.
To analyze high speed circuit, no value should be set on area and design should
run on high clock frequency. The area is increased by 70% when the design is
synthesized for maximum speed and speed is increased nearly thrice.
37
6.2.2 Area of different modules in the design
The area occupied by different components in the design are calculated and
tabulated in table 7. The values are taken from the same scenario as specified in
6.2.1.2 .We can see that most of the area in the design is consumed by the
multipliers. There are four multipliers in the whole design, which have different
word lengths and different areas. The DFFs are used to analyze the critical path
delay in the design. The input MUX is used to select one of the inputs depending
on the quadrant. There is one 14- bit multiplier (squarer in Fig. 16) and 19-bit
multiplier to calculate the product of first two sub functions. These two multipliers
plays a major role in area occupancy and hence the number is bigger for the S1, S2
Function module in table 7. The S3 Function also consists of one Multiplier and
two LUTs to store the coefficients. A 19-bit multiplier gives the final result, which
is a product of three sub functions.
Table 7 .The area of different modules in the design
Module name
Input DFF
Output DFF
Input MUX
S1,S2 Function
S3 Function
Multiplier
Other cells
Total Area
Area
147
200
60
5432
2402
3148
22
11400
Units
6.2.3 Maximum clock frequency
The speed of the design is measured in terms of maximum clock frequency, which
is the sum of delays in critical path. The maximum clock frequency is calculated
by the minimum time period required for the design, which we can see in table 5.
The maximum clock frequency is the inverse of the total critical path time, which
is calculated as below.
The critical path of the design will be affected if there is any change in the
operating voltage, temperature and threshold voltage. Maximum speed of the
design is calculated by setting a high frequency clock value as a constraint to the
design without setting any value for area. The Synthesis scenario is specified in
6.2.1.1.
1
1
189
5.29
38
6.3 Timing and Area analysis
6.3.1 Timing
The design is synthesized with Synopsys design compiler using both 65nm LPHVT
and LPLVT CMOS libraries. The timing behavior is analyzed under different
scenario like operating the design at different supply voltage. Fig. 19 shows
graphical description for critical path delay in both LPHVT and LPLVT cases and
it is clear that the delay is almost double in case of LPHVT compared to LPLVT .
Table 8. Timing values for LPHVT and LPLVT
Clock(nS)
Voltage
1.1
1.15
1.25
1.3
LPHVT (ns)
18.4
17.0
14.3
13.2
LPLVT (ns)
10.0
9.5
8.78
8.4
20
18
16
14
12
10
8
6
4
2
0
LPHVt
LPLVt
1,1
1,15
1,25
1,3
Voltage(V)
Figure 19. Timing behavior for LPHVT and LPLVT
6.3.2 Area
The Parabolic Synthesis methodology architecture is designed focusing on the area
occupied on hardware. The major components in the design, which occupies the
area are multipliers and the optimizations in the design, which are explained
chapter 5.3.3, reduce the area of the design. Fig. 20 shows the graphical
description of area, which is analyzed under different supply voltages i.e. 1.1V,
1.15V, 1.25V and 1.3V for both LPHVT and LPLVT libraries.
39
Table 9 . Area for LPHVT and LPLVT
Voltage
1.1
1.15
1.25
1.3
LPHVT (
11338
11331
11327
11317
)
LPLVT (
11470
11580
11743
11820
)
The values in table 9 are calculated for minimum area design by keeping
set_max_area to 0 during the synthesis. From Fig. 20, there is a slight decrement
in area of the design as the supply voltage is increased in case of LPLVT ,where as
it is opposite in case of LPHVT .
11900
11800
11700
Area(µm2)
11600
11500
11400
LPHVt
11300
LPLVt
11200
11100
11000
1,1
1,15
1,25
1,3
Voltage(V)
Figure 20. Area analysis for LPHVT and LPLVT
6.4 Power analysis
6.4.1 Dynamic Power
The dynamic power depends on the switching activities in the design, capacitance,
operating frequency and voltage as equated in (6.1).
40
(6.1)
where α = Switching activity (0-1, 1-0)
c = node capacitance
f = clock frequency
v = supply voltage
The power consumption for the design in case of both LPHVT and LPLVT is
analyzed using the Primetime tool. The total power is divided into static power
and dynamic power and the major contribution for dynamic power is switching
activities of the design. The netlist file generated after synthesis is used in this
process and the switching activities are generated as a Value Change Dump
(VCD) file in the post synthesis verification process. The dynamic power
consumed by the design, for different voltages in both cases is plotted in Fig. 21
and tabulated in table 10 and 11.
Table 10. Power results (LPHVt)
Voltage(V)
1.1
1.15
1.25
1.3
Cell internal
Power(µW)
70.26
(46.5%)
77.03
(46.6%)
91.9
(46.7%)
101.4
(46.9%)
Net switching
power(µW)
74.11
(49.2%)
81.23
(49.1%)
96.4
(48.9%)
105.3
(48.7%)
Leakage
power(µW)
6.5
(4.3%)
7.2
(4.3%)
8.67
(4.4%)
9.4
(4.4%)
Total
Power(µW)
150.9
(100%)
165.5
(100%)
197.0
(100%)
216.2
(100%)
Table 11. Power calculation (LPLVt)
Voltage(V)
1.1
1.15
1.25
1.3
Cell internal
Power(µW)
87.3
(13.9%)
101.8
(14%)
131.4
(13.4%)
147.3
(13%)
Net switching
power(µW)
84.3
(13.4%)
86.29
(11.8%)
104
(10.6%)
113.5
(10%)
41
Leakage
power(µW)
450
(72.7%)
540
(74.2%)
740
(76%)
870
(77%)
Total
Power(µW)
626.7
(100%)
730
(100%)
983
(100%)
1133
(100%)
The dynamic power calculated is plotted in Fig. 21 and dynamic power is more in
case of LPLVt compared to LPHVt..
Dynamic power(µW)
300
250
200
150
LPHVt
100
LPLVt
50
0
1,1
1,15
1,25
1,3
supply voltage (V)
Figure 21. The dynamic power consumption (LPHVT &LPLVT)
6.4.2 Static Power
The leakage power is mainly depends on transistor width, threshold voltage and
zero threshold leakage current as shown in (6.2).
.
6.2 The leakage power at different supply voltage for both LPHVT and LPLVT is
analyzed and plotted in Fig.22 and 23. The leakage power is more in case of
LPLVT compared with LPHVT.
42
Leakage power(µW)
10
9
8
7
6
5
4
3
2
1
0
1,1
1,15
1,25
1,3
Voltage(V)
Leakage power(µW)
Figure 22. The Leakage power (LPHVT)
1000
900
800
700
600
500
400
300
200
100
0
1,1
1,15
1,25
1,3
Voltage(V)
Figure 23. The Leakage power (LPLVT)
The static powers in case of both technology libraries are plotted in Fig.22 and 23.
The static power is very less in case of LPHVT where it is a major contribution in
total power in case of LPLVT. In case of LPLVT, there is rapid growth in static
power consumption as the supply voltage is increasing.
43
6.5 Energy
The power dissipation is the main objective when analyzing the efficiency of a
design. The main source of power includes the switching activities, leakage
current and short circuit current. The switching activities are dynamic in nature
and leakage current depends on the threshold voltage. To analyze the efficiency of
the design, energy consumed by the design is considered as a measure as
calculated according to (6.3).
∙
6.3
The design is analyzed for energy consumption in case of both LPHVT and LPLVT
libraries. The values in table 12 are measured at the maximum frequency of the
design. The energy consumption is less in case of LPLVT compared to LPHVT.
Table 12. Energy consumption in the design
Voltage
1.1
1.15
1.25
1.3
Energy consumption(nJ)LPHVT
2.65
2.67
2.68
2.71
Energy consumption(nJ)LPLVT
1.71
1.78
2.06
2.18
3,00E‐09
Energy consumption(J)
2,50E‐09
2,00E‐09
1,50E‐09
LPHVT
1,00E‐09
LPLVT
5,00E‐10
0,00E+00
1,1
1,15
1,25
1,3
Voltage(V)
Figure 24 . Energy consumption in the design
44
From Fig. 24, the energy consumption is less in case LPLVT compared to LPLVT.
Since the leakage power is more in LPLVT, it is not considered to be good
approach to proceed to LPLVT libraries. A tradeoff is applied when considering
the libraries when synthesizing the design.
6.6 Comparison
The sine function approximation using Parabolic Synthesis combined with linear
interpolation results are compared with other implementations. In this chapter, the
differences between the implementations and comparing the static and dynamic
behavior of the designs.We considered the implementation of exponential function
using Parabolic Synthesis methodology [13] and compared with the sine function
approximation design. The Parabolic Synthesis methodology provides a feasibility
to implement any unary function by using the same architecture with different
coefficients. Peyman implemented the exponential function using Parabolic
Synthesis using four sub functions in order to get more than 15 bits resolution. The
same accuracy is achieved in our implementation with two sub functions and
linear interpolation. The CORDIC algorithm is used to calculate the
Trigonometric, and the Logarithmic functions, where the application demands
simple hardware and low precision as well. The CORDIC algorithm only consists
of adders, shifters and a lookup table and hence it occupies less hardware area
[19].The algorithm approximately increase the precision by 1 bit per iteration,
which means at least 15 iterations are needed achieve a 15 bits precision[19].
Table 13. ASIC Synthesis analysis
Methodology
Area(
CORDIC
Parabolic Synthesis(4 Stage)
Parabolic Synthesis(2 stage) with
Interpolation
)
19150
26400
19400
maximum operating
frequency(
)
11.5
48
55
Table 14. Total Dynamic power analysis
Methodology
CORDIC
Parabolic
Synthesis(4
stage)
Parabolic
Synthesis(2
stage) with
Interpolation
Cell internal
power (µW)
132.76
54%
162.04
51%
Switching power
(µW)
114.61
46%
158.57
49%
Leakage
power(nW)
64.71
76.99
49.8%
77.57
50.1%
23.78
45
97.78
The area of the design with four sub functions is high since third and fourth sub
function requires more computations. The CORDIC algorithm occupies less area
but high latency makes it to use in low speed applications. The power
consumption is less in CORDIC when compared with Parabolic Synthesis
Methodology and higher when maximum frequency of operation is considered
[13] [19].The Parabolic Synthesis combined with linear interpolations is designed
to calculate the values for all the four quadrants using the same interpolation
coefficients, which is also an added advantage with the design and causes that its
area increases slightly in relation to the other implementations.
Table 13 and 14 shows the comparison of both the implementations in different
aspects. The maximum operating frequency for the design is imported from table
8, the inverse of 18.4 ns. We cannot compare both the designs in numbers since
the switching activities used to analyze the designs may not be same and hence we
compared in sense of percentage of the total power.
46
CHAPTER
7
7 Conclusion
The approximation of the sine function is successfully implemented using the
Parabolic Synthesis methodology combined with a linear interpolation technique.
An accuracy of more than 15 bits is achieved with hardware implementation of the
design. Hardware optimization techniques, which replace multipliers with an
MCM unit, removal of adder result lesser area and power in the design. The design
is synthesized in different scenario to estimate both dynamic and static behavior
and results are compared with the CORDIC implementation [13] [16]. It can be
concluded that the Parabolic Synthesis Methodology is efficient only in
computational speed with almost 4.8 times faster. Parabolic Synthesis shows high
degree of parallelism, which results shorter critical path in the design. Parabolic
Synthesis methodology has flexibility to change the design to implement for other
unary functions by modifying the set of coefficients.
47
CHAPTER
8
8 Future work
The linear interpolation technique combined with the parabolic synthesis
methodology can be used for implementing other unary functions as well. Sine
function can also be approximated by replacing the linear interpolation technique
with other interpolation methods. The accuracy may be increased by
approximating sine function with second-order interpolation technique.
48
9
CHAPTER
9 References
[1] Erik Hertz, “Parabolic Synthesis”, Licentiate thesis in engineering
[2] P.T.P Tang, Table-lookup algorithms for elementary functions and their error
analysis, Mathematics and Computer science division, Argonne National
Laboratory.
[3] www.wikipedia.org/wiki/Polynomial_interpolation
[4] Chi-Chou Kao, High-performance CORDIC rotation algorithm based on lookahead techniques International Journal of Electronics Vol. 98, No. 8, August
2011, 1075–1089
[5] Erik Hertz and Peter Nilsson, "A Methodology for Parabolic Synthesis," a
book chapter in VLSI, In-Tech, Vienna, Austria,ISBN 978-3-902613-50-9, pp.
199-220, September 2009.
[6] Volder J. E. The CORDIC trigonometric computing technique. In IRE Trans.
Electronic Computing, volume EC-8, pages 330 - 334, 1959.
[7] Venkatesh Jakke, Low Leakage Digital Design, a CORDIC Processor, Master
of Science thesis.
[8] http://spiral.net/index.html
[9] http://en.wikipedia.org/wiki/Linear_interpolation
[10] www.wikipedia.org/wiki/Interpolation#Piecewiseconstant_interpolation
[11] http://en.wikipedia.org/wiki/Interpolation#Polynomial_interpolation
[12] Design of FFTs using CORDIC and Parabolic Synthesis for the twiddle
factors by Muhammed Waqas Shafiq Nauman Hafeez
49
[13] Peyman Pouyan, Erik Hertz, and Peter Nilsson, “A VLSI Implementation of
Logarithmic and Exponential Functions Using a Novel Parabolic Synthesis
Methodology Compared to the CORDIC Algorithm,” in the Proceedings of the
European Conference on Circuit Theory and Design (ECCTD 2011), August 2931, 2011, pp. 729-732, Linkoping, Sweden
[14] www.ece.neu.edu/students/sbabaeiz/paper_poster/interpolation.htm
[15] http://www.mathsisfun.com/gradient.html
[16] Master‘s Thesis Report, A VLSI Implementation of Logarithmic and
Exponential Functions Using a Parabolic Synthesis Methodology, which is
Compared to the CORDIC Algorithm By Peyman Pouyan
[17] A. Volnei A. Pedroni, Digital Electronics and Design with VHDL.
[18] http://repositories.lib.utexas.edu/bitstream/handle/2152/1472/arbaughj
20424. Pdf.
[19] Orri Tomasson, Implementation elementary functions for a fixed point
SIMD DSP Coprocessor, Examensarbete, Linkopoing.
50
10
CHAPTER
10 List of Figures
Figure 1. The Method of interpolation -------------------------------------------------- 14 Figure 2. The Error analysis for 1 sub function & linear interpolation ------------- 16 Figure 3. The Error analysis for 2 sub functions & linear interpolation ------------ 17 Figure 4. The Normalization for Sine function----------------------------------------- 19 Figure 5. The Developing the first sub function --------------------------------------- 20 Figure 6. fhelp (x) & second sub function s2(x) --------------------------------------- 21 Figure 7. K values for the third sub function ------------------------------------------- 22 Figure 8.Normalization of original angle ----------------------------------------------- 22 Figure 9. The Design flow ---------------------------------------------------------------- 24 Figure 10. The Top level architecture --------------------------------------------------- 25 Figure 11. The Architecture of 2 Stage Parabolic Synthesis ------------------------- 26 Figure 12. The Multiple Constant Multiplier (MCM) unit --------------------------- 27 Figure 13. The architecture of the third sub function (Linear interpolation) ------- 28 Figure 14. The Addsub Component ----------------------------------------------------- 29 Figure 15. Two's complement module -------------------------------------------------- 29 Figure 16. Block diagram with wordlengths ------------------------------------------- 30 Figure 17. The Design Test setup -------------------------------------------------------- 35 Figure 18. The absolute error of the original function compared with the result
using 1, 2, and 3 sub functions ----------------------------------------------------------- 36 Figure 19. Timing behavior for LPHVT and LPLVT ---------------------------------- 39 Figure 20. Area analysis for LPHVT and LPLVT -------------------------------------- 40 Figure 21. The dynamic power consumption (LPHVT &LPLVT) ------------------- 42 Figure 22. The Leakage power (LPHVT) ----------------------------------------------- 43 Figure 23. The Leakage power (LPLVT) ------------------------------------------------ 43 Figure 24 . Energy consumption in the design ----------------------------------------- 44 51
CHAPTER
11
11 List of Acronyms
CORDIC
ASIC
MCM
LPLVT
LPHVT
VHDL
MSB
VCD
CMOS
RTL
SDF
Coordinate Rotation Digital Computer Algorithm
Application Specific Integrated Circuit
Multiple Constant Multiplication
Low Power Low Threshold Voltage
Low Power High Threshold Voltage
Very High Speed Integrated Circuits Hardware Description
Language
Most Significant Bit
Value Change Dump
Complementary Metal Oxide Semiconductor
Register Transfer Level
Standard Delay Format
52
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement