2013 5th International Conference on Computer Science and Information Technology (CSIT) ISBN: 978-1-4673-5825-5

Projective Rational Arithmetic with Floating Point

Vaclav Skala
Department of Computer Science and Engineering
Faculty of Applied Sciences, University of West Bohemia
CZ 306 14 Plzen, Czech Republic
http://www.VaclavSkala.eu
AbstractNumerical computation is used nearly in all software
packages. Today, computation is based on the floating point
representation IEEE754 which offers single and double precision
while quadruple or extended precision are supported very rarely
in current programming languages. In this paper an alternative
representation is presented which offers a higher range of digits
of a fraction and a higher range of exponents. It is based on the
projective extension of the Euclidean space. Due to this
representation the division operation can be nearly eliminated,
that leads to faster and more robust computation. The presented
projective rational arithmetic is especially convenient for vectorvector architectures like GPU.

observed. One typical example is a matrix multiplication in the

case of large matrices, when Strassen algorithm is used to
speed-up the computation, but the error of computation is
growing fast.
TABLE I.

Keywords numerical computation; floating point; IEEE754;

precision of computation; linear algebra; computer graphics;
scientific computation; projective geometry; GPU computation.

I.

INTRODUCTION

Numerical computation is a kernel of computational

packages in nearly all fields of applied computer science.
Todays systems are processing larger and larger data sets,
problems solved are more and more converging to ill
conditioned numerical problems, computation is far from being
precise and numerical representation is limited to integers and
floating number representation with a limited precision using
mostly IEEE754 standard. Numerical instability or imprecision
caused several very expensive catastrophes, e.g. well known
sinking of the Sleipner oil platform, Ariane spaceship failure
[13] [15] etc.
In the old ages Algol68 programming language [22], [24]
had the long construction, which enabled to extend
numerical precision, e.g.
long long real pi = 3.14159 26535 89793 23846 26433
83279 50288 41971 69399 37510;
Algol68 had a significant influence to many programming
languages including C, resp. C++ and it was used to write an
operation system ICL VME, too. A good summary of the Algol
68 features is described in [24]. Several modifications,
incomplete implementations and new implementations have
been reported [23]. Unfortunately the long construction is not
available in current programming languages and current
floating representations are mostly limited to single or double
precision; very rarely quad or extended precision is supported.
Double precision seems to be sufficient, however especially in
scientific computing, geometry computations etc. the precision
is not acceptable and unreliability of computations is very often

260

FLOATING POINTS BIT STRUCTURE

Name

Digits

E min

E max

B 16

Half

10+1

14

15

B 32

Single

23+1

126

127

B 64

Double

52+1

1022

1023

B 128

Quad

112+1

16382

16383

The current IEEE 754 specification is highly optimized

from the technology or hardware point of view. However
there is a question whether this approach is viable in future.
There are some implementations aiming exact computation,
e.g. [4], but the computation is just slow. On the other hand in
practice we usually accept a limited precision if the precision is
high enough and the results have accuracy required.
In this paper an alternative approach using projective
extension of the Euclidean representation will be presented and
discussed.
II.

ALTERNATIVE REPRESENTATIONS

The problem with a precision can be solved by using exact

arithmetic approach [4], but computational time is high in
general. There are other alternative ways how to increase
a precision of computation as in the majority of cases we do
not need exact solution having very long fractional part. High
attention given especially to:

rational arithmetic based on fractions with integer

numbers, i.e. a rational number is expressed as
where

are integers

continuous fractions, e.g.

978-1-4673-5825-5/13/$31.00 (c) 2013 IEEE

2013 5th International Conference on Computer Science and Information Technology (CSIT) ISBN: 978-1-4673-5825-5

logarithmic representation in which multiplication is

becoming a summation, division is a subtraction etc.
However summation itself is complicated requiring high
memory capacity [1].

However there are several drawbacks if rational arithmetic with

integers or continuous fractions are used:

current processors are optimized to the IEEE754 floating

point representation

division operation is slow and leading to inaccuracy and

to high energy consumption

The precision of computation can be estimated using interval

arithmetic. If the x representation, resp. y is in the
interval
, resp.
then the precision of basic
operations can be estimated as follows:

IV.

BASIC OPERATIONS USING PROJECTIVE

Robustness of computation is a key issue in many

sophisticated computational algorithms as sophisticated
engineering problems solved today might be ill conditioned.
Let us explore basic properties of the projective representation
from a numerical precision and robustness point of view.
However different data structures have to be considered, i.e.

projective scalar

projective vector

etc.

Fundamental operations with:

A. scalars

addition, resp. subtraction

multiplication

division

B. vectors
It can be proved that the division operation is causing
substantial inaccuracy of computation.
There is a question whether the projective extension of the
Euclidean representation can bring advantages over the
classical using the floating point representation single or
double precision.
III.

addition resp. subtraction

scalar multiplication (dot product)

vector multiplication (cross product)

PROJECTIVE REPRESENTATION

Projective extension of the Euclidean space is usually

considered only as a tool useful for geometric
transformations in computer graphics and computer vision.
A point
is represented in projective space by
homogeneous coordinates
, where:
is the
homogeneous coordinate [2] [7] [10] [17] [18] [20]. The
transformation from the projective space to the Euclidean space
is given as
and
In mathematics a different notation, more convenient, is used:

It can be seen that the division operation can be hidden to

the homogeneous coordinate , resp. . As the direct result
of that is that the division operation is not needed, higher
precision can be expected. However, it should be noted that
problems with exponent overflow or underflow must be
considered.
Using projective representation and principle of duality [3]
[5] [6], several algorithms are simpler and easy to implement as
well [8] [11] [16] [19] [21].

Note that the projective vector is different from a vector which

consists of projective scalars.
C. exponent normalization
Exponents due to arithmetic operations tend to grow or
become smaller. It means that the exponent overflow or
underflow is to be check and exponents can be normalized.
This is very simple operation as it means that the same value in
the exponent is to be added or subtracted from both
numerator and denominator as well.
Extraction of an exponent for a single and double precision
is defined as
EXP := (FP_value land MASK) shr m;
where: land is bitwise and operation, shr is shift right, MASK
is the binary mask and m is the argument for the shift
operation. In this way we can manipulate with the exponent,
i.e. read or rewrite it. The sequence is quite simple; special
cases are not considered. The corresponding values for
different precision are defined in Tab.2.

261

2013 5th International Conference on Computer Science and Information Technology (CSIT) ISBN: 978-1-4673-5825-5

TABLE II.

Disadvantages

HARDWARE CONSTATNTS

Precision

MASK

Exp_Digits

Single

&7FC0

255

Double

&7F80

2047

D. Comparison operation

Current hardware does not support projective rational

floating point, but the additional computational cost of that
is low, but should be considered

Operations are approx. two times longer if not vectorvector architecture or SSE instructions are used

Value of exponents have to be controlled there is a

possibility of exponent overflow or underflow, but easily
solved by addition or subtraction to numerator and
denominator in hardware.

The comparison operation is a little bit tricky as the

condition

Normalization can be made in software without a

significant slowdown of computations.

i.e.

Projective scalars have to have homogeneous coordinate nonnegative, i.e.

and

TABLE III.

The condition is to be replaced as follows:

V.

ADVANTAGES OF PROJECTIVE RATIONALS

There are some advantages and disadvantages of the

proposed projective representation. Let us explore some of
them for better understanding of a projective rational arithmetic
application potential.
Advantages

The mantissa is actually doubled due to the hidden

division operation by the homogeneous value a there is a
higher range of the fractional part

The exponent range is higher. If a single precision is used,

the range is
to
, i.e.the range is actually

The division operation is eliminated by multiplication of a

homogeneous value in which denominator is hidden

Infinity can be handled properly, i.e. division by a value

close or equal to zero does not cause floating point
overflow

262

If double precision for numerator and denominator is used

is used, actually a quadruple extended precision is
implemented; if quadruple precision is available we get
more that 2 times better precision
Simple implementation on vector-vector architectures, like
GPU available projective Library P-Lib [25]

REPRESENTATION OF VALUES

Vector of projective scalars

or the values must be multiplied by the sign of the

homogeneous coordinate of the value. In practice this is not a
complication as the condition is easy to implement in hardware
for general case, if not limited to positive homogeneous value
in general.
It should be noted that projective representation is used in
computer graphics and computer vision naturally, not only for
geometrical operations, but also for intersection computations
etc.

There is a significant difference between vector of

projective scalar values and projective vector, i.e.

Projective vector

means equal projectively

and users have to be careful in the mathematical and
expression formulations.

In the case of iterative methods on the current CPU longer

computation time is to be expected. The given approach is
not convenient for application of iterative methods on
CPU due to exponent values control in software
VI.

EXPERIMENTAL VERIFICATION

Let us consider following simple examples for

demonstration of the proposed projective rational arithmetic in
linear algebra.
Gauss elimination
Gaussian elimination method is well known for solving a
system of linear equations
where:

It generally produces an upper triangular matrix, row

echolon form, and then solves the unknown in the backward
run.
The structure of the Gaussian elimination method is
for k := 1 to n
for i := k+1 to m
for j := k+1 to m

2013 5th International Conference on Computer Science and Information Technology (CSIT) ISBN: 978-1-4673-5825-5
The expression for

can be rewritten as

As the value of
can be very small, i.e.
, and
division by a denominator could cause significant inaccuracy
or floating point overflow, exchange of rows is made in
practice. If the projective notation is used
where:

is the homogeneous part of the expression. If

and
is the reciprocal value on ,
then we can write
It should be noted that a reciprocal value
is
actually a swap of
and
values. The scalar value
represented in the projective notation is
, i.e. the
homogeneous value is non-zero, usually . It means that no
division operation is needed, however it is still hidden in the
homogeneous coordinate and pivoting has to be used.
As the solution of a system linear equations
is
equivalent to extended cross-product [9] [12] [14] we can write

point representation. From the precision point of view, it

offers higher range of exponents and also significantly wider
range for a fraction representation.
The presented approach is convenient for vector-vector
hardware architectures including GPU. If used on CPU with
SSE instructions it is slightly slower than the computation
with the Euclidean notation, but offers higher precision
natively. If double representation is used it offers more than
quadruple or extended representation. It means that in several
engineering computations it can be used to significantly
increase precision of computation. Current C# implementation
of the P-Lib library [25] uses current GPU hardware and it
was intended for geometrical applications in E3. In the future
work the P-Lib library will be modified for numerical
computations in general and optimized.
ACKNOWLEDGMENT
The author would like to thank to anonymous reviewers for
their critical comments and hints that improved this paper
significantly. Thanks also belong to students and colleagues at
the University of West Bohemia, Plzen and VSB-Technical
University, Ostrava for their suggestions. The project was
supported by the Ministry of Education of the Czech Republic,
project No. LH12181.

where
REFERENCES
[1]

and

[2]

where is the solution of the linear system of equations in the

projective form, i.e.
.

[3]
[4]

It can be seen that no division is required unless we need to

express the computed value in the Euclidean space.

[5]

As direct consequence of the equivalence we can easily

solved many computational problems without the division
operation.
The projective rational arithmetic with floating point has
been verified experimentally for stability and precision of
computation and inversion of the Hilbert matrix was used,
which converge to a singular matrix with the growing size. The
experiments proved the expected properties of the proposed
approach, details can be found in [12]. [25].

[6]
[7]

[8]

[9]

[10]

VII. CONCLUSION
Projective rational arithmetic with floating point was
described and fundamental arithmetic operations were
described. The projective representation using homogeneous
coordinates is used in computer graphics and computer vision
and its application enabled to solve many problems in more
effective way. As it was shown the projective representation is
convenient for general numerical computation as well as it has
several advantages of the standard single or double floating

[11]

[12]

[13]

Arnold,M.G: A VLIW Architecture for Logarithmic Arithmetic,

Proceedings of the Euromicro Symposium on Digital System Design
(DSD03), IEEE, 2003
Bloomenthal,J., Rokne,J.: Homogeneous Coordinates, The Visual
Computer, Vol.11,No.1, pp.15-26, 1994.
Coxeter,H.S.M.: Introduction to Geometry, John Wiley, 1969.
Fogel,E., Halperin,D., Wein,R.: CGAL Arrangements and Their
Applicaqtions A Step by Step, Springer Verlag, 2012
Hartley,R, Zisserman,A.: MultiView Geometry in Computer Vision,
Cambridge Univ. Press, 2000.
Johnson,M.: Proof by Duality: or the Discovery of New Theorems,
Mathematics Today, December 1996.
Skala,V.: A New Approach to Line and Line Segment Clipping in
Homogeneous Coordinates, The Visual Computer, Vol.21, No.11,
pp.905 914, Springer Verlag, 2005
Skala,V.: Length, Area and Volume Computation in Homogeneous
Coordinates, International Journal of Image and Graphics, Vol.6., No.4,
pp.625-639, 2006
Skala,V.: Barycentric Coordinates Computation in Homogeneous
Coordinates, Computers & Graphics, Elsevier, Vol. 32, No.1, pp.120127, 2008
Skala,V.: Intersection Computation in Projective Space using
Homogeneous Coordinates, Int.Journal on Image and Graphics, Vol.8,
No.4, pp.615-628, 2008
Skala,V.: Geometry, Duality and Robust Computing in Engineering,
WSEAS Trans.on Computers, Vol.11, No.9, ISSN 1109-2742, E-ISSN
2224-2872, pp.275-291, 2012
Skala,V: Duality and Intersection Computation in Projective Space
with GPU Support, WSEAS Trans.on Mathematics,
Vol.9.No.6.pp.407-416, 2010
Skala,V.: Duality and Intersection Computation in Projective Space
with GPU support, Applied Mathematics, Simulation and Modeling ASM 2010 conference, NAUN, Corfu, Greece, pp.66-71, 2010

263

2013 5th International Conference on Computer Science and Information Technology (CSIT) ISBN: 978-1-4673-5825-5
[14] Skala,V.: Duality and Robust Computation, ICCOMP12, 16th
WSEAS Int.Conf. on Computers ICCOMP12, Kos, Greece, pp.172-117,
2012
[15] Skala,V.: Projective Geometry and Duality for Graphics, Games and
Visualization, Course SIGGRAPH Asia 2012, Singapore,
ISBN 978-1-4503-1757-3, 2012
[16] Skala,V.: S-Clip E2: A New Concept of Clipping Algorithms,
SIGGRAPH Asia Posters, accepted for publication, 2012
[17] Skala,V., Kaiser,J., Ondracka,V.: Library for Computation in the
Projective Space, 6th Int.Conf. Aplimat, Bratislava, pp. 125-130, 2007
[18] Stolfi,J: Oriented Projective Geometry, Academic Press, 2001.
[19] Thomas,F., Torras,C.: A Projective invariant intersection test for
polyhedral, The Visual Computer, Vol.18, No.1, pp.405-414, 2002
[20] Yamaguchi,F.: Computer-Aided Geometric Design A Totally Four
Dimensional Approach, Springer Verlag, 1996
[21] Yamaguchi,F., Niizeki,M.: Some Basic Geometric Test Conditions in
Terms of Plcker Coordinates and Plcker Coefficients, The Visual
Computer, Vol.13, pp.29-41, 1997

can be implemented in Cg/HLSL on GPU as follows:

WEB references
[22] Algol 68 - http://en.wikipedia.org/wiki/ALGOL_68
#Operating_Systems_written_in_ALGOL_68
[23] Algol 68- download, http://algol68.sourceforge.net/
[24] Algol68 summary, http://en.wikipedia.org/wiki/ALGOL_68
[25] C# Library for computation in the projective space P-Lib, download
http://www.kiv.zcu.cz/vyzkum/software/index.html

APPENDIX A
The cross product in 4D defined as

float4 cross_4D(float4 x1, float4 x2, float4 x3)

{
float4 a;
a.x=dot(x1.yzw, cross(x2.yzw, x3.yzw));
a.y=-dot(x1.xzw, cross(x2.xzw, x3.xzw));
// or a.y=dot(x1.xzw, cross(x3.xzw, x2.xzw));
a.z=dot(x1.xyw, cross(x2.xyw, x3.xyw));
a.w=-dot(x1.xyz, cross(x2.xyz, x3.xyz));
// or a.w=dot(x1.xyz, cross(x3.xyz, x2.xyz));
return a;

or more compactly
float4 cross_4D(float4 x1, float4 x2, float4 x3)
{
return ( dot(x1.yzw, cross(x2.yzw, x3.yzw)),
-dot(x1.xzw, cross(x2.xzw, x3.xzw)),
dot(x1.xyw, cross(x2.xyw, x3.xyw)),
-dot(x1.xyz, cross(x2.xyz, x3.xyz)) );
}

264