0% found this document useful (0 votes)
54 views75 pages

Using The Gaussian Elimination Method For Large Banded Matrix Equations

This document describes methods for solving large banded matrix equations using Gaussian elimination. It presents an original "thrifty" method that allows solving such equations on computers with limited memory by systematically swapping matrix contents between RAM and hard disk. This avoids inefficiencies of conventional virtual memory and reduces disk read/write operations. Equations are constructed from a 2D Laplace equation and solved using a full Gaussian elimination solver and the new thrifty method. Results are compared to the known analytical solution.

Uploaded by

Lucas Santos
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
54 views75 pages

Using The Gaussian Elimination Method For Large Banded Matrix Equations

This document describes methods for solving large banded matrix equations using Gaussian elimination. It presents an original "thrifty" method that allows solving such equations on computers with limited memory by systematically swapping matrix contents between RAM and hard disk. This avoids inefficiencies of conventional virtual memory and reduces disk read/write operations. Equations are constructed from a 2D Laplace equation and solved using a full Gaussian elimination solver and the new thrifty method. Results are compared to the known analytical solution.

Uploaded by

Lucas Santos
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 75

Using the Gaussian Elimination Method for

Large Banded Matrix Equations

Jerome P.-Y. Maa


School of Marine Science
Virginia Institute of Marine Science
College of William and Mary
Gloucester Point, VA 23062

Ming-Hokng Maa
Massachusetts Institute of Technology
Cambridge, MASS

Changqing Li
Andersen Consulting LLP
Minneapolis, MN

Qing He
East China Normal University
ShangHai, China.

Special Scientific Report No. 135


VIMS
SH
1
V48
no.135 January 1997

'Jl\\~S A.RCH\\/ES OCT 1 1997'


Table of content

Table of Content. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . I

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . I I

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Methodology- . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

Simultaneous Linear Algebra Equations ......•........•.•..... 4

General Banded Matrix Equation ............................•• 6


Full Banded-matrix Equation Solver ..................•....... 8
A Thrifty Banded Matrix Equation Solver ...•...........•..... 9
Discussion and Conclusions . . . • . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

Appendix I, source Listing of Program P_BANDED.FOR ..••.... I-1

Appendix II, Source Listing of Program P_THRIFT.FOR •.....• II-1


Appendix III, Source Listing of FORTRAN Program ID.FOR III-1

Appendix IV, Source Listing of MATLAB Program P_PLOT.M IV-1


Appendix v, source Listing of P EXACT.M ..... . .... .. ... .. .. V-1

I
Abstract

A special book-keeping method was developed to allow

computers with limited random access memory but sufficient hard-

disk space to feasibly solve large banded matrix equations by


using the Gaussian Elimination method with Partial Pivoting. The

computation time for this method is excellent because only a

minimum number of disk readjwrite is required, freeing processor

time for number crunching.

II
INTRODUCTION

Numerical solutions of elliptic partial differential

equations with high resolution or large domains are typically


approached through iteration methods, e.g., Successive-Over-

Relaxation (SOR) (Roache 1972), Multi-Grid (MG) (Brandt 1984),


etc. This is primarily because direct approaches such as the
Gaussian Elimination method cannot feasibly solve large sets of

linear algebra equations (or matrix equations) with limited

computer memory. For example, a square two dimensional domain

with 200 grids per side will generate a banded coefficient matrix
with a dimension.of 400 x 40000. If four bytes are used to

represent real numbers, a computer would require approximately 64

megabytes (MB) of memory to store this matrix alone. If 16 bytes

are used to represent complex numbers with improved accuracy, the

memory requirement could easily increase to 256 MB. This memory

requirement is typically too large for general computers.

using the Gaussian Elimination method for solving a partial

differential equation, however, has two advantages: (1) The

numerical scheme is simple and has been available for more than

20 years; (2) The computation time is not affected by the type


and complexity of boundary conditions. In contrast, the

convergence rate for iteration methods typically degrades with

Neumenn type boundary conditions or complex boundary geometries.

In order to practically implement the Gaussian Elimination

method, however, the enormous memory requirements must be

addressed. With today's advances in hard-disk storage, computer

1
hard-disks on the order of gigabytes(GB) or more have become

standard and commonplace. It has also become a standard practice

for many operating systems to automatically satisfy an

application's memory requirements with virtual memory, swapping

random access memory with hard-disk storage. This built-in

virtual memory, however, has very low efficiency because it was

never designed for computations with matrices of very large

orders. In this study, a method for solving large banded matrix

equations by systematically swapping the contents of a high order

matrix between memory and hard-disk is presented. This report

will detail the construction of the banded matrix equation, and

compare the original Gaussian Elimination method of solution,

versus the thrifty banded matrix solver method of solution.

Computer source codes are listed in the Appendices and are also

available on disk for registered user.

METHODOLOGY

A simple two dimensional elliptic Laplace equation (Eq. 1}

is used to demonstrate the construction of a banded matrix

equation.

~:~ + ~~~ = 0 ••••••••••••••••••••••••••••••••••••••••••• ( 1)

For simplicity, we are using Dirichlet boundary conditions in

this study. The boundary conditions are specified at the border

of a square domain, from 0 to n, are as follows: p = 2 at x = o;


P = 1 at y = o; and p =o at both x = y = n. For this example,

2
we can find the analytical solution to compare the results
obtained from each approach.

Using an s-transform on the y component yields the


analytical solution:

p = ~~ 2
{[1-2(-1) 0 ]cosh(nx) - [ [1-2(-1)"]coth(nrr) +
1 nrr

csch(nrr) ] sinh(nx) + 1} sin(ny) ......... (2)

The analytical solution (with n=4000) is plotted in Fig. 1. The


MATLAB program that calculated Eq. 2 can be found in Appendix v.
Notice that the analytical solution does not exactly meet the

boundary condition ¢ = 1 at y = 0. This is because the solution

is in the form of Fourier series, which converges to the average,

i.e., zero, at the boundary y = o. In the neighborhood of this

boundary, however, ¢ increases sharply and approaches 1 when y >


1xlo-3 • Two selected profiles of the analytical solution

demonstrate this situation in Fig. 2. In Fig. 2a, a linear scale

on the y-axis indicates that p ~ 1 in the neighborhood of y = o.

In Fig. 2b, a semi-log scale on the y-axis indicates that P = 0

at y = o, but p increases very fast to approach 1 when y > 0.001.

The contour plot in Fig. 1 was generated with a spatial

resolution of ~x = ~y = 0.03141592, which is not high enough to

show the significant change of p for 0 < y < 0.001. Thus the p

values at y = 0.001 instead of y = 0 were used with other p

values to construct the contour plot (Fig. 1).

3
SIMULTANEOUS LINEAR ALGEBRA EQUATIONS

Using the central finite difference method, Eq. 1 can be


rewritten as

p.l",J
1 . + r p .. _ - 2(1+r) P,· J. + r P,· J"+l + P,-+ 1, j. = 0
l,J 1 I I
••••••••••• (3)

where r = (~xj~y) 2 , and ~x and ~y are the grid sizes selected in


the x and y directions, respectively. This equation is typically
referred to as a five-point-approximation because only five
points are involved in one iteration loop of an iteration method
(e.g., SOR or MG) or to build a band matrix equation.
To demonstrate the construction of the banded matrix e
points in both the x and y directions, is initially chosen. See

2 + p11 = 0
2 + P,z = o
2 + P,3 = o

2 + Ps 4 p9 + P,o + P19 = 0
2 + p9 4 P,o + 0 + Pzo = 0
P, + 1 4 p11 + P12 + P21 = 0
Pz + P,, 4 P,z + P13 + P22 = 0
p3 + P12 4 p13 + P14 + p23 = 0 ................. (4)

P79 + Psa - 4 Ps9 + P9o + P99 = 0


Pao + Ps9 - 4 P9o + 0 + P1oo = 0

Psa + P97 - 4 P9s + P99 + O = 0


Ps9 + P9s - 4 P99 + P1oo + 0 = 0
P9o + P99 - 4 p 100 + 0 + 0 = o

4
Fig. 3. The solutions are the p values at each intersection
point. In Fig. 3 , they are marked as p 1 , p 2 , p 3 , ••• , p 99 , p 100 •

Application of Eq. 3 at these intersection points (from p 1


to quation, a course grid, ~x = ~y = 0.28559933, i.e., 12 grid
p 100 ) generates the following simultaneous linear equations.

The constants in the above equations can be moved to the

righthand side, and the simultaneous linear equations can be

written as a band matrix equation:

A P = B • • • • • • •• • • • • • • • • • • • • • • • • • • • • • • •• • • •• • • • • • • • • • • • • • • (5)

where A is a banded coefficient matrix, P is a column matrix that

contains the unknown variable p in the study domain, and B is a

column matrix that contains the given boundary conditions.

0 0 0 0 1 1· 1 1 1 1 1
0 0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0 0
0 1 1 1 1 1 1 1 1 1 1
A = -4 -4 -4 -4 -4 -4 -4 -4 -4 -4 -4 ........ ( 6)
1 1 1 1 1 1 1 1 1 1 0
0 0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0 0
1 1 1 1 1 1 1 1 0 0 0

5
BT = [- 3 -2 -2 •• - 2 - 1 0 0 .. . 0 1 0 .• . 0 1 0 . . . 0 0] . • (8 )

Matrix A has a dimension of M x N, (21 X 100) 1 where M is

the band width and N is the total number of unknown points within

the study domain. The band width is the sum of the upper band

width, MU, the lower band width, ML, and the unit width of
diagonal elements, i.e., M = MU + ML + 1. The upper band width,

MU, is the maximum count of the element P.1+ 1 IJ. from the diagonal

element, P; j. Here the count represents the number of grid


1

points between P 1..J and P 1.+ 1


I I
J.. The measurement begins in the i th

column and j+1 th row and continues with increasing row number

until the maximum row number or the upper boundary is reached,

and then continues on the first grid point (or the lower

boundary) on the i+1_th column, and ends at the j_th row.

Similarly, ML is the maximum count between element P.1" 1 1



j
and P 1...
11

GENERAL BANDED MATRIX EQUATION

In general, the coefficient matrix, A, can be written as

follows. Note that in Eq. 2, r need not be 1. Also in Eq. 8,

6
0 0 0 0 a1,u a1,n-L a1, n-L-1 a1,n-1 al,n
0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0
0 ad-1, 2 ad-1,3 ad-1, u-1 ad-1, u ad-1,n-L ad-1, n-L-1 ad-1, n -1 ad-1, n

1l = ad, 1 ad,2 ad,3 ad,u-1 ad,u ad, n-L ad, n-L-1 ad,n- 1 ad,n .. (9)

d+1, 1 ad+1, 2 ad+1,3 ad•1,u-1 ad•l, u ad+1, n-L ad+1, n-L-1 ad+1,n-1 0
0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0
am, 1 am,2 am,3 am,u-1 a
m,u
a
m, n-L 0 0 0

a1,u'' a1,u+1' , and a 1,n are not necessary all in the same row,

i.e., row 1. In this example, however, they are all in the same

row because of the simple boundary geometry. If the boundary

geometry was irregular, then these elements would not be in the

same row. Similarly, am, 1 , am,Z' ... , and am,L are not necessary all

in row M.

A problem arises when the above stated approach is used to

solve an elliptic partial differential equation when the required

resolution is high, or the study domain is large, i.e., the

banded matrix equation is large. To illustrate this problem, the

previous example will be solved by both the original Gaussian

Elimination method with Partial Pivoting and the thrifty banded

matrix solver developed for this study.

7
FULL BANDED MATRIX EQUATION SOLVER

To solve Eq. 1 using the full banded matrix, A, a program

P_BANDED.FOR (see Appendix I), which use the FORTRAN subroutines

(CGBFA and CGBSL) from Linpack (Dongarra et al. 1979), was


established. This program performs adequately on a personal

computer (PC) with the coarse grid. Fig. 4 plots the results

using a MATLAB program, P_PLOT.M (see Appendix IV), which reads

the output generated by P_BANDED.FOR directly. From Fig. 4, it

is obviously to see that a higher resolution is required. When

the grid size is increased to 102 x 102, however, the program

P_BANDED.FOR would require approximately 42 MB of memory to

execute. While a Windows 3.1/95 based PC would use virtual

memory to execute the program, the computing time would be very

long because Windows would be constantly swapping between memory

and hard-disk. Because of this reason, we changed to a Unix

workstation SUN SPARC II computer with 32 MB of memory to do the

job. It is still very slow for this size because of disk and

memory swap. But we eventually got the results, which is

identical to Fig. 1.
The input data for the program P_BANDED.FOR was generated

using another FORTRAN program, ID.FOR, given in Appendix III.


The input file generated by ID.FOR was as general as possible so

users could easily modify the input file for different equations.

Notice, however, that this program was only programmed for a

Dirichlet boundary value problem. For Neumann type boundary

conditions, the routine that formulates the banded coefficient

8
matrix would be quite different. The solution procedure for the
banded matrix equation, however, would remain essentially the

same.

A THRIFTY BANDED MATRIX EQUATION SOLVER

Notice that there are many zero elements in A, especially when

the band width is large. Eq. 2 indicates that only the diagonal

terms (ad,i with j=1, N), the two immediate off-diagonal terms (ad-
. and ad+ • with j=1, N), the upper border terms (a 1 • with j=1,
1, J 1, J ,J

N) , and the lower border terms ( am,j with j=1, N) are possibly

non-zero terms. All other elements are zero. The entire banded

matrix A, therefore, can be stored in two much smaller matrices,

v and T, each with a dimension of N x 5. Matrix V stores the

five rows that contain possible non-zero elements as v.1, k with k =


1 to 5 and i = 1 to N. Their locations (all integer values) in

the original banded matrix A are stored in matrix T. In this


way, A can be stored more compactly. For comparison, the band

matrix presented in the introduction would only require 1.2 MB as

opposed to the original 64 MB of memory to store.

In solving Eq. 2, the two small matrices, v and T, are

recalled one block at a time to construct a working matrix, w,


with dimensions of (M+ML) x (NK+M), where NK is determined by the

available computer memory.

9
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0

0 0 0 0 0 0 0 0

w= 0 0 0 0 0 0 0 0 ( 1 0)

ad,1

d •1,1 ad•1,nk•m

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0
am,1
am,u am,u•l am, n+m-1

Notice that in the working matrix, the original banded

matrix was.shifted ML rows downward. This is just following the

approached used in Linpack. Also notice that some data were

excluded after column NK+M when fetch a block (length NK) of data

from V and T to construct W. These data will be recovered in the

construction of next W to maintain data integrity.

10
0 0 0 0 0 u1 ,M U1,NK+M

0 0 0 0 U2, M-1 U2,M U2,NK+M

0 0 0 u3, M-2 uJ, M-1 u3,M U 3,NK+M

0 0 0
0 0 0
0 0 0
0 0 UM-2,3

W= 0 UM-1,2 UM-1,3 UM-1 ,M-2 UM-1 ,M-1 UM-1,M UM-1,NK+M


. (11)

~.1 ~.2 ~.3 ~,M-2 ~,M-1 ~,M d M,NK+M

~+1,1 ~+1,2
m
M+1,3 mM+1 ,M-2 mM+1 ,M-1 ~+l,M m
M+1,NK+M

~·2,1 ~+2,2 ~2,3 ~2,M-2 mM+2 ,M-1 ~-2,M ffiM+2,NK+M

~ML-2,1 ~+ML-2,2 ~+ML-2 ,3 ~+ ML-2, NK+M

mM+ML-1,1 ~+ML-1,2 ~+ML-1 ,3 ~+ML-1,NK+M

~+ML, 1 Jt..ML,2 ~ML, 3 ~+HL,NK+M

In general, the larger the NK, the smaller the number of disk

IfO, thereby reducing the processing time, however, this speed


reduction is only marginally because the number of disk I/O is

already minimized in this process. As a rule of thumb, NK should

be at least 4 x M.
After applying the procedure of Gaussian Elimination Method

for a banded matrix on w (Dongarra et al. 1979), the results were

saved in the first M rows (row 1 toM). The space from row M+1

to M+ML is used to save the multipliers. This is also the

typical approach used in Linpack. In this study, however, only

11
the first M row, from column 1 to NK, was written into hard disk
as a temporary file. The multipliers in row M+1 to M+ML were

ignored because thery is no use for a single column matrix B.

Only for those cases that matrix B is a rectangular matrix, we


then need to store these multipliers.
After write a disk file, data in the block from column NK+1
to NK+M is then moved to the beginning of the working matrix,

i.e., column 1 to column M. The data ignored when expending the

first block of V and T into W is then retrieved, and the second

block (also lenth of NK) of V and T is then expanded into the

working matrix. Some data at the end of this block must again be

ignored due to a lack of space in W. This process repeats until


the entire V and T have been processed. During this forward

elimination process, many temporary files will have been written

into the hard-disk. These files are written in binary form to

improve I/O times and to conserve disk space.

The backward substitution method implemented in the FORTRAN

subroutine CGBSL from Linpack is then used to find the solution,

P, also one block at a time. This routine is relatively simple:

the last disk file was read first (and deleted), and the backward

substitution routine is used to calculate the solutions for this

block. This process also repeated for all the blocks to find out
P.

Using the finite difference method with dx = dy = 0.031105,

i.e., a grid size of 102 x 102. With the thrifty banded matrix

solver, we only need a working matrix that has a size of 310 x

12
400. Thus, the size of executable codes reduces to about 2 MB.
Of course, one needs about 40 MB of disk space for writing

temporary files, which were deleted at the end. Under this


condition, the example problem given previously can be solved

using a personal computer with reasonable time. For example, the


computer time for solve the same problem using a Pentium 75 MHz

PC was about 160 sec. The results are identical to that given in

Fig. 1.
Source list of this thrifty banded matrix solver, a FORTRAN

program, P_THRIFT.FOR, is given in Appendix II. The input data

file is the same as that used for the program P_BANDED.FOR, and

the results were also plotted using the MATLAB program P_PLOT.M.

Some of the subroutines (e.g., DOMAIN, ICAMAX, CAXPY, CSCAL,

IDISTR, and IDISTL) used in this program are identical to those

used in the program P BANDED.FOR. But they are put together to

have a complete file.

DISCUSSION AND CONCLUSIONS


The Gaussian Elimination method involves two steps: forward

elimination and backward substitution. Mathews (1987) pointed

out that the number of computations involved in the forward

elimination is proportional to N3 , where N is the order of the

coefficient matrix. In contrast, the number of computations

involved in backward substitution is proportional to N2 •

computation time, therefore, is mostly spent in forward

elimination. If there are many linear systems (all with the same

13
coefficient matrix) to be solved,. then the forward elimination
should only be calculated once, and the results, multipliers,

should be saved for different B matrices. While the book-keeping


procedure developed in this study could easily be adapted to
solve multiple linear systems, the details are out of the scope

of this study.
For very large banded matrix equations, i.e., a large N, the

round-off error may become significant. To improve accuracy, it

may become necessary to use double precision (8 bytes for real

numbers). This increase in precision, however, does not


significantly increase the memory requirement because the size of

the executable code is small to begin with, and sufficient hard-

disk space is not difficult to obtain.

While only a simple application was presented in this note,

this method can be used in many applications, especially when

there are Neumenn boundary conditions and an irregular boundary


geometry. For those situations, the formulation of the banded

matrix equation may differ, but the banded matrix solver remains

the same.
The use of the Gaussian Elimination Method with Partial

Pivoting, the development of a special book-keeping procedure,

and the high availability of large hard-disks allow the practical

solution of large banded matrix equations. The feasibility of

efficiently calculating the solution of an elliptic partial

differential equation with large study domain andjor high

resolution on a personal computer was demonstrated. While

14
; .
developed for computers with limited resources, however, this

method could easily compute domains on a global scale when

processed on computers with large memory and/or disk resources.

ACKNOWLEDGMENTS

Financial support from the u.s. Department of the Interior,

Minerals Management Service, Office of International and Marine

Minerals, under contract number 14-35-0001-30740 is sincerely

acknowledged.

REFERENCES

Brandt, A., 1984, Multigrid Techniques: 1984 Guide with

Applications to Fluid Dynamics, GMD study, number 85, Postfach

1240, Birlinghoven, D-5205 St. Augustin 1, Germany.

Dongarra, J.J., J.R. Bunch, C.B. Moler, and G.W. Stewwart, 1979,

Linpack Users Guide, SIAM.

Mathews, J. H., 1987, Numerical Methods for Computer Science,

Engineering, and Mathematics, Prentice-Hall, Inc., Englewood

Cliffs, New Jersey, 507pp.

Roache, P. J., 1972, Computational Fluid Dynamics, Hermosa

Publishers, Albuquerque, New Mexico, 434 pp.

15
2
1.8 1.6
1

1.2 0.8

1 2 2.5 3
X

Fig. 1. Contours Plot of the Analytical Solution.

16
a n=10000
1.5

at x=0.5 ..,..

·--.,.............
p
···,"-....
·--...................____
0.5 at x=2.0--~·--........
.... -··-·-·-..........,
--....
-·--··-·-·---.........._..__
··.......• ..
o+---~--~--~--~----.---.---.----.---.--~
0 rt/5 3rt/5 4rt/5

2~-----------------------------------------,

1.5 n=10000 at x=0.;_:.5'--.-

p
. ····~-----·----·-···· ...•

at x=2.0 ~--.....
\,
0.5
\\
\
'
O~TT~=-~onmr-rrnnm-.llnmr-rr~rr-TTTmm-.-~mm
1 E-06 1 E-05 0.0001 0.001 0.01 0.1 1 10
y

Fig. 2. Analytical Solution Profiles at Selected x locations.

17
12 I T
I I

11
P.n I Pwc

10
pg I I I pgg

9
Pa I
8
I I I I I
p6
7
y Ps I
6

5
p4 I
4
p3 I
p2 P,2 P92
3

P, P, p31 Pa, Pg,


2

1
I I
I
1 2 3 4 5 6 7 8 9 10 11 12

Fig. 3. Coarse Grid Points in the Study Domain.

18
y

1 2 3
X

Fig. 4. Numerical Solution Using the Coarse Grid

19
Appendix I. Source Lists of FORTRAN program P_BANDED.FOR
program p_banded
c
c This program is to calculate the velocity potential (p) in the
c area 0.0 <= x <= 3.1416, 0.0 <= y <= 3.1416.
c
c a thrify storage method was constructed first, and then,
c convert to that required by LINPACK, and uses a minor modified
c LINPAK subroutines to solve it.
c
c Because of the formidible slow pace when runing this program
c with high resolution, we can only limited the parameter KQ to a
c small number.
c
parameter (iq=110,jq=110, IW=310, kq=400, Lq=3, mq=203)
implicit complex (z)
character*45 title, label
character*20 infile, outfile, recfile
character*1 id, q1
common /matrx/ za(kq,S), ia(kq,S), zb(kq)
common /gener/ mp, np, dx, dy, r, p(iq,jq), id(iq,jq)
common /unknw/ n,imap(kq) ,jmap(kq) ,irow(iq)
common /labbl/ ind(iq) ,jin(iq,Lq) ,jout(iq,Lq)
common /works/ zabd(iw,kq), ipvt(kq)
c
c boundary conditions
c
bcx0=2
bcxn=O
bcy0=1
bcyn=O
c
print*,'Select 1. p_s.grd; 2. p_m.grd 3. p_L.grd: '
read(*,*) iption
go to (10, 12, 14), iption
10 infile='p s.grd'
outfile='P bs.out'
recfile='p=bs.rec'
go to 20
12 infile='p_m.grd'
outfile='p_bm.out'
recfile='p_bm.rec'
go to 20
14 infile='p_L.grd'
outfile='p_bL.out'
recfile='p_bL.rec'
20 continue
c
open(9, file=infile, status='old')
open(10, file=outfile, status='unknown')
open(11, file=recfile, status='unknown')

I-1
c
c mp,np : Max. grid numbers in x andy direction, respectively.
c dx,dy : grid sizes in x andy direction, respectively.
c
read(9,' (a45) ') title
write(11,' (a45) ') title
read(9,*) mp,np,dx,dy
if(mp .gt. iq) then
print*, 'IQ is smaller than MP, Change IQ to ',mp
stop
end if
if(np .gt. jq) then
print*,'JQ is smaller than nP, Change JQ to' ,np
stop
end if
c
r == (dx/dy)**2
c
c ID code for each grid point
c 0, interior point
c
c 1, upper boundary condition
c 2, lower boundary condition
c 3, left boundary condition
c 4, right boundary condition
c
c 5, left bottom corner B.C.
c 6, left top corner B.C.
c 7, right bottom corner B.C.
c 8, right top corner B.C.
c
c e, grid point that is not included in the study domain
c
c Th~ following is an example
c
c j
c "' np 6111111111118eeeeeeeeeeeeeeeeeee
c 3000000000004eeeeeeeeeeeeeeeeeee
c 30000000000001111111111111111118
c 30000000000000000~00000000000004
c 30000000000000000000000000000004
c 30000000000000000000000000000004
c 30000000000000000000000000000004
c 52200000000000000000000000222227
c eee30000000000000000000004eeeeee
c eee30000000000000000000004eeeeee
c eee30000000000000000000004eeeeee
c 1 eee52222222222222222222227eeeeee
c 1 mp
c
c
----------------------------------------------> i
read(9,30) label

I-2
write(11,30) label
read(9,30) label
write(11,30) label
30 format (a50)
do j=1,np
jj=np-j+1
read(9,45) jns, (id(i,jj) ,i=1,mp)
45 format(i5, (110a1) )
if(jns .ne. jj) then
write(11,50) jj,jns
50 format(' Sequence is wrong for ID input at' ,2i5)
stop
end if
end do
print*,'Completed reading ID code matrix'
close(9)
c
c construct the unknow COLUMN matrix X, in the matrix eq. AX=B
c find each unknown's location: imap(map) ,jmap(map)
c
c ind(i} : no. of isolated sector in each culomn, x grid.
c for this particular case, ind(i) are all 1.
c because no land points in the middle of study domain.
c jin(i,index) begin grid number for an isolated sector in a
c column
c jout(i,index): end grid number for an isolated sector in a
c culomn
c map the number of total unknow, or the length of X.
c later, it is reassigned as N
c irow(i) the total number of unknow vel. Potential in each
c column
c
map=O
do i=1,mp
icount=O
in=O
iout=1
index=1
do j=1,np
if(id(i,j) .eq. '0' then
icount=icount+1
map=map+1
imap(map)=i
jmap(map)=j
if(in .eq. 0) then
jin(i,index)=j
in=1
iout=O
end if
end if
c
if( id(i,j) .eq. '1' .and. iout .eq. 0) then

I-3
jout(i,index)=j
iout=1
in=O
index=index+1
end if
end do
irow(i)=icount
ind(i)=index-1
c
c ind(i) should be >= 1, except for the entire column are all
c boundary points. if not, something wrong.
c
write(11,55) i, ind(i), irow(i), jin(i,1), jout(i,1)
55 format(' i,ind,irow,jin,jout=' ,SiS)

if( ind(i) .gt. Lq) then


write(11,60) i, ind(i), Lq, ind(i)
60 format(' At i=' ,i4,' Ind(i)=' ,i2,' > Lq=' ,i2/
* ' Change Lq to' ,i3,' and re-run')
stop
end if
end do
c
c n: length of the banded matrix
c
n=map
if(n .lt. O.S*kq) then
write(*,65) n, kq
65 format ('
-------------------------------------------------'/
* 'N=' ,iS,' << KQ=' ,iS//' It is better to reduce KQ,',
* ' length of the Banded Matrix,'/' So that KQ is not',
* ' >>than that required.'/' Instead, one should',
* ' increase the length of the working matrix, JW,'/
* ' in order to reduce disk I/O and computing time.'/
* ' You may continue, or re-run< c/r > : ')
read ( *, ' ( a1) ' ) q1
if( q1 .eq. 'r' .or. q1 .eq. 'R') stop
end if
if(n .gt. kq) then
write(11,70) n, kq, n
70
format('--------------------------------------------------'1
* N=' ,iS,' > KQ=' ,iS//
I

* ' Increases KQ t o ' ,iS, ' and re-run')


stop
end if
c
c Set up the two small matrices ZA and IA for storing the banded
c matrix and find out the band width, work size, etc.
c
mu=O
I-4
mL=O
write(11,80)
80 format (' i j ia ( 1) , .... , ia(4),ia(5)
* ' za ( 1) . . . za ( 5) zb')
do map=1,n
i=imap(map)
j=jmap(map)
c
call domain(i, j, map, ierr)
if( ierr .eq. 0) then
write(11,90)i,j,ia(map,1) ,ia(map,2) ,ia(map,3)·,ia(map,4),
* ia(map,5), real(za(map,1)) ,real(za(map,2)),
* real(za(map,3)), real(za(map,4)) ,real(za(map,5)),
* real ( zb (map) )
90 format(1x,2i5,2x,5i5,2x,6f6.2)
else
print*,'stop, error in DOMAIN at i,j,map=', i,j,map
print*,'Check the recording file for details.'
stop
end if

if( ia(map,5) .ne. 0 ) then


mu_c=abs(ia(map,5) - ia(map,3)
else
mu c=O
end if
if(mu .lt. mu c) mu=mu c
if( ia(map,1) .ne. 0 ) then
mL c=abs(ia(map,1) - ia(map,3)
else -
· mL c=O
end if
if(mL .lt. mL_c) mL=mL c
end do
c
Lda=2*mL+mu+1
m=mL+mu+1
write(*,100) m, n, mu, mL, Lda
100 format(' Band width, m = ',i7/' Data point, N = ',i7/
* 'Upper B.W., mu = ',i7/' Lower B.W., mL = ',i7/
* ' Lda for LINPAK = ',i7)
c
if(m .gt. mq) then
print*,'M > MQ, please increase MQ t o ' ,m
stop
end if
c
c change the thrifty storage to band storage specified by LINPAK,
c LINPAK uses rows ML+1 through 2*ML+MU+1 of ZABD to store the
c original banded matrix. In addition, the first ML rows in ZABD
c are used for elements generated during the triangularization.
c The total number of rows needed in zABD is 2*ML+MU+1 .
I-5
C The ML+MU by ML+MU upper left triangle and the ML by ML lower
c right triangle are not referenced.
c
c but first clear the matrix
c
do map==1,n
do k==1,Lda
zabd(k,map) == (0.0, 0.0)
end do
end do
c
c position the diagonal element
c
kkk==mu+1+mL
do map==1,n
if( ia(map,3) .eq. map) then
zabd(kkk,map)== za(map,3)
else
print*,'Check conversion,map,za(i,3)==' ,map,za(map,3)
stop
end if
c
c position the low triangular elements
c
do k==1,2
if(ia(map,k) .gt. 0) then
ik==kkk + ia(map,3) - ia(map,k)
jk==map - (ia(map,3) - ia(map,k)
zabd(ik,jk)==za(map,k)
end if
end do
do k==4,5
if(ia(map,k) .gt. 0) then
ik==kkk + ia(map,3) - ia(map,k)
jk==map + ia(map,k) - ia(map,3)
zabd(ik,jk)==za(map,k)
end if
end do
end do
c
print*,'Completed conversion to LINPACK standard'
print*,'The matrix length is from 1 to' ,n
print*,' Select a checking domain from n1 ,n2 =='
read(*,*) n1, n2
do map==n1,n2 .
write(*,130) imap(map), jmap(map), map, real(zb(map)),
* (real(zabd(k,map)) ,k==1,lda)
130 format(' i,j,map,b==' ,2i5,i8,f8.1/(10f7.1))
read ( * , ' ( a1) ' ) q1
end do
c
c call gauss elimination to slove the complex band matrix equ.
I-6
c .; .

print*,'Calling CGBFA, wait


call CGBFA(Lda, n, ml, mu, info)
c
if (info .eq. 0) then
print*,'Successfull called CGBFA'
print*,'Want to store results in a file <y/n>
read ( * , ' ( a1) ' ) q1
if (q1 .eq. 'y' .or. q1 .eq. 'Y') then
write(11,132)
132 format(' i zabd(i,j), j=1,n')
do i=1,Lda
write(11,13S) i, (real(zabd(i,j)) ,j=1,n)
13S format(iS/(10f8.3))
end do
end if
else
write(*,140) info
140 format(' error code (info) =', iS,' After called CGBFA')
end if
c
call CGBSL(lda, n, ml, mu, zB)
c
if (info .eq. 0) then
print*,'Successfully called CGBRL'
else
write(*,160) info _,
160 format(' error code (info) - , iS,' after called CGBFA')
end if
c
do map=1, n
i=imap(map)
j=jmap(map)
p(i,j)=real(zb(map)
end do
do j=1,np
p(1,j) = bcxO
p(mp,j) = bcxn
end do
do i=1,mp
p(i,1) = bcyO
p(i,np) = bcyn
end do
p(1,1) = O.S*(bcyO+bcxO)
p(1,np) = O.S*(bcyn+bcxO)
p(mp,1) = O.S*(bcyO+bcxn)
p(mp,np) = O.S*(bcyn+bcxn)

write(10,' (a45) ') title


write(10,' (2i5,2f12.8) ') mp,np,dx,dy

I-7
do j=1,np
write ( 1 0 , 2 0 0 ) j , (p ( i , j ) , i = 1, mp )
200 format(i5/(10f8.4))
end do·

close(10)
close ( 11)

print *,'program stop'


stop
end
c
c--------------------------------------------------------
c
subroutine domain(i, j, map, ier)
c
parameter (iq=110,jq=110, IW=310, kq=400, Lq=3)
implicit complex (z)
character*1 id
common /matrx/ za(kq,S), ia(kq,S), zb(kq)
common /gener/ mp, np, dx, dy, r, p(iq,jq), id(iq,jq)
common /unknw/ n,imap(kq) ,jmap(kq) ,irow(iq)
common /labbl/ ind(iq) ,jin(iq,Lq) ,jout(iq,Lq)
common /works/ zabd(iw,kq), ipvt(kq)
common /tempo/ zam(203,kq)
c
dimension iid(9)
c
c The finite difference governing equation, the Laplace equation,
c with r = (dx/dy)**2, is
c
c p(i-1,j) +r*p(i,j-1) -(2+2r)*p(i,j) +r*p(i,j+1) +p(i+1,j) = o
c
c In storage, the coefficient of p(i-1,j) is stored in (map 1 1) 1
c the coefficient of p(ilj-1) is stored in (map 1 2) 1
c the coefficient of p(ilj) is stored in (map 1 3) ,
c the coefficient of p(ilj+1) is stored in (map 1 4) 1
c and the coefficient of p(i+11j) is stored in (map 1 5)
c
ier=O
do k=1 1 9
iid(k)=O
end do
icount=O
c
c For those grid point that are not neighbored to a boundary
c
if(id(i-1,j) .eq. '0' .and. id(i+1 1 j) . eq. ' o ' . and .
* id(i 1 j-1) .eq. 1 0' .and. id(i 1 j+1) . eq. 1 0 ' ) then
za(map,1)=1.0
za(map,2)=r
za(map 13)=-2.0*(1.0+r)

I-8
za(map 1 4)=r
za(map 1 5)=1.0
ia(mapll)=map-idistl(i j)+1 1

ia(map 1 2)=map-l
ia(map 1 3)=map
ia(mapl4)=map+1
ia(mapiS)=map+idistr(ilj)-1
zb(map)=(0.0 1 0.0)
icount=icount+1
iid(1)=1
end if
c
c for those has a Boundary grind point on the left
c
if(id(i-1 1 j) .eq. 1
3 1
.and. id(i 1 j-1) .eq. 1
01 .and.
* id ( i I j +1) . eq. 1 0 1 ) then
za(map 1)=0.0
1

za(map 1 2)=r
za(map 1 3)=-2.0*(1.0+r)
za(map 1 4)=r
za(map 1 5)=1.0
ia(map 1 1)=0
ia(map 1 2)=map-1
ia(map 1 3)=map
ia(map 1 4)=map+1
ia(map 1 5)=map+idistr(i 1 j)-1
zb(map)=(-2.0 1 0.0)
icount=icount+1
iid(2)=1
end if
c
c for those has a Boundary grind point on the right
c
if(id(i+1 1 j) .eq. 4 1
.and. id(i 1 j-1)
1
.eq. 1
01 .and.
* id ( i j +1) . eq.
I 0
1
then
1
)

za(map 1 1)=1.0
za(map 1 2)=r
za(map 1 3)=-2.0*(1.0+r)
za(map 1 4)=r
za(map 1 5)=0.0
ia(map 1 1)=map-idistl(i 1 j)+1
ia(map 1 2)=map-1
ia(map 1 3)=map
ia(map 1 4)=map+1
ia(map 1 5)=0
zb(map)=(O.O, 0.0)
icount=icount+1
iid(3)=1
end if
c
c for those has a Boundary grind point on the top
c

I-9
if(id(i,j+1) .eq. '1' .and. id(i-1,j) .eq. '0' .and.
* id(i+1,j) .eq. '0') then
za(map,1)=1.0
za(map,2)=r
za(map,3)=-2.0*(1.0+r)
za(map,4)=0
za(map,5)=1.0
ia(map,1)=map-idistl(i,j)+1
ia(map,2)=map-1
. ia (map, 3) =map
ia(map,4)=0
ia(map,S)=map+idistr(i,j)-1
zb(map)=(O.O, 0.0)
icount=icount+1
iid(4)=1
end if
c
c for these grid points that has a Boundary at bottom
c
if(id(i,j-1) .eq. '2' .and. id(i-1,j) .eq. '0' .and.
* id(i+1,j) .eq. '0' ) then
za(map,1)=1.0
za(map,2)=0
za(map,3)=-2.0*(1.0+r)
za(map,4)=r
za(map,5)=1.0
ia(map,1)=map-idistl(i,j)+1
ia(map,2)=0
ia(map,3)=map
ia(map,4)=map+1
ia(map,S)=map+idistr(i,j)-1
zb(map)=(-1.0, 0.0)
icount=icount+1
iid(5)=1
end if
c
c For those grid points that are left bottom corner
c
if(id(i-1,j) .eq. '3' .and. id(i,j-1) .eq. '2') then
za(map,1)=0.0
za(map,2)=0
za(map,3)=-2.0*(1.0+r)
za(map,4)=r
za(map,5)=1.0
ia(map,1)=0
ia(map,2)=0
ia(map,3)=map
ia(map,4)=map+1 .
ia(map,S)=map+idistr(i,j)-1
zb(map)=(-3.0, 0.0)
icount=icount+1
iid(6)=1

I-10
end if
c
c For those grid points that are left top corner Boundary points
c
if(id(i-l,j) .eq. '3' .and. id(i,j+l) .eq. '1') then
za(map,l)=O.O
za(map,2)=r
za(map,3)=-2.0*(1.0+r)
za(map,4)=0.0
za(map,S)=l.O
ia(map,l)=O
ia(map,2)=map-l
ia(map,3)=map
ia(map,4)=0
ia(map,S)=map+idistr(i,j)-1
zb (map) = (- 2 . 0, 0 . 0)
icount=icount+l
iid(7)=1
end if
c
c For those grid points that are right bottom boundary points
c
if(id(i+l,j) .eq. '4' .and. id(i,j-1) .eq. ' 2 ' ) then
za(map,l)=l.O
za(map,2)=0
za(map,3)=-2.0*(1.0+r)
za(map,4)=r
za(map,S)=O.O
ia(map,l)=map-idistl(i,j)+l
ia(map,2)=0
ia· (map, 3) =map
ia(map,4)=map+l
ia(map,S)=O
zb(map)={-1.0, 0.0)
icount=icount+l
iid(8)=1
end if
c
c For those grid points that are right top boundary points
c
if(id(i+l,j) .eq. '4' .and. id(i,j+l) .eq. '1') then
za(map,l)=l.O
za(map,2)=r
za(map,3)=-2.0*(l.O+r)
za(map,4)=0
za(map,S)=O.O
ia(map,l)=map-idistl(i,j)+l
ia(map,2)=map-l
ia(map,3)=map
ia(map,4)=0
ia(map,S)=O
zb(map)=(O.O, 0.0)

I-ll
icount=icount+1
iid(9)=1
end if
c
if( icount .ne. 1) then
ier=1
write(11,10) ier, icount, (iid(k) ,k=1,9)
10 format(' ier, icount=' ,2i5,' id(1)-id(9)=' ,9i2)
end if
return
end
c
c----------------------------------------------------------------
c
integer function idistr(i,j)
parameter (iq=110,jq=110, IW=310, kq=400, Lq=3)
implicit complex (z)
character*1 id
common /matrx/ za(kq,S), ia(kq,S), zb(kq)
common /gener/ mp, np, dx, dy, r, p(iq,jq), id(iq,jq)
common /unknw/ n,imap(kq) ,jmap(kq) ,irow(iq)
common /labbl/ ind(iq) ,jin(iq,Lq) ,jout(iq,Lq)
common /works/ zabd(iw,kq), ipvt(kq)
c
c Calculates the distance between points (i+1,j) and (i,j). It
c means how many grid points, which are solved for, are in
c between. It counts vertically from the point(i,j) and up
c (i,j+1),(i,j+2), ... , (i,np), then (i+1,1), (i+1,2), ... , t o
c (i+1,j). In the matrix equation, this distance represents the
c up band width at that particular diagonal element (i,j).
c
do k=1,ind(i)
if(j .ge. jin(i,k) .and. j .le. jout(i,k) ) mark1=k
end do
c
do k=1,ind(i+1)
if(j .ge. jin(i+1,k) .and. j .le. jout(i+1,k) ) mark2=k
end do
c
idd=jout(i,mark1)-j+1 + j-jin(i+1,mark2)
c
do k=mark1+1, ind(i)
idd=idd + jout(i,k)-jin(i,k)+1
end do
c
do k=1,mark2-1
idd=idd + jout(i+1,k)-jin(i+1,k)+1
end do
c
idistr=idd
c
return

I-12
end
c
c
integer function idistl(i,j)
c
parameter (iq=110,jq=110, IW=310, kq=400, Lq=3)
implicit complex (z)
character*1 id
common /matrx/ za(kq,S), ia(kq,S), zb(kq)
common /gener/ mp, np, dx, dy, r, p(iq,jq), id(iq,jq)
common /unknw/ n,imap(kq) ,jmap(kq) ,irow(iq)
common /labbl/ ind(iq) ,jin(iq,Lq) ,jout(iq,Lq)
common /works/ zabd(iw,kq), ipvt(kq)
common /tempo/ zam(203,kq)
c
c Calculate the distance between point (i-1,j) and (i,j). The
c distance means how many grid points, which are solved for, are
c in between. It counts vertically from (i,j) downward (i,j-1),
c (i, j -2), ... , (i,1) and restarted from previous i line
c (i-1, np), (i-1, np-1), ... , to (i-1, j).
c
do k=1,ind(i-1)
if(j .ge. jin(i-1,k) .and. j .le. jout(i-1,k)) mark1=k
end do
c
do k=1,ind(i)
if(j .ge. jin(i,k) .and. j .le. jout(i,k)) mark2=k
end do
c
idd=jout(i-1,mark1)-j+1 + j-jin(i,mark2)
c
do k=markl+1,ind(i-1)
idd=idd+jout(i-1,k)-jin(i-1,k)+1
end do
c
do k=1,mark2-1
idd=idd+jout(i,k)-jin(i,k)+1
end do
c
idistl=idd
return
end
c
c
SUBROUTINE CGBFA(LDA, N, ML, MU, INFO)
c
c Original Date of WRITTEN Aug. 14 1978 by MOLER, C. B.,
c PURPOSE Factors a COMPLEX band matrix by Gaussian elimination
c and partial pivoting.
c
c CGBFA is usually called by CGBCO, but it can be called
c directly with a saving in time if RCOND is not needed.

I-13
c
c On Entry
c
C ABD COMPLEX(LDA, N), contains the matrix in band storage. The
c columns of the matrix are stored in the columns of ABD and
c the diagonals of the matrix are stored in rows ML+1
c through 2*ML+MU+1 of ABD. See comments below for details.
c
c LDA INTEGER, the leading dimension of the array ABD
c LDA must be .GE. 2*ML + MU + 1 .
c
c N INTEGER, the order of the original matrix.
c
c ML INTEGER, number of diagonals below the main diagonal.
c 0 .LE. ML .LT. N
c
c MU INTEGER, number of diagonals above the main diagonal.
c 0 . LE. MU . LT. N. More efficient. if ML . LE. MU
c
c On Return
c
c ABD an upper triangular matrix in band storage and the
c multipliers which were used to obtain it. The
c factorization can be written A = L*U where L is a
c product of permutation and unit lower triangular matrices
c and U is upper triangular.
c
c IPVT INTEGER(N), an integer vector of pivot indices.
c
c INFO INTEGER
c = 0 normal value.
c = K if U(K,K) .EQ. 0.0 . This is not an error condition,
c but it does indicate that CGBSL will divide by zero if
c called. Use RCOND in CGBCO for a reliable indication
c of singularity.
c
c This uses rows ML+1 through 2*ML+MU+1 of ABD. In addition, the
c first ML rows in ABD are used for elements generated during
c the triangularization. The total no. of rows needed in ABD is
c 2*ML+MU+1. The ML+MU by ML+MU upper left triangle and the ML
c x ML lower right triangle are not referenced.
c
c LINPACK. This version dated 08/14/78 .
c Cleve Moler, University of New Mexico, Argonne National Lab.
c
c Subroutines and Functions called from this subroutine
c
c BLAS CAXPY, CSCAL, ICAMAX
c Fortran ABS,AIMAG,MAXO,MINO,REAL
c
c REF: DONGARRA J.J., BUNCH J.R., MOLER C.B., STEWART G.W.,
c *LINPACK USERS GUIDE*, SIAM, 1979.

I-14
C ROUTINES CALLED CAXPY,CSCAL,ICAMAX
c
c Revised by Jerome P.-Y. Maa on Nov. 96.
c 1. Move the matrix abd(Lda, 1) and Ipvt(1) to common block. so
c that the limitiation of LDA = IW can be relaxed. In other
c words, the parameter IW specified don't have to be exactly
c equal to LDA.
c 2. uses implicit statement for complex variables
c
parameter (IW=310, kq=400)
implicit complex (z)
character*1 q1
common /works/ zabd(iw,kq), ipvt(kq)
REAL CABS1
CABS1(ZDUM) = ABS(REAL(ZDUM)) + ABS(AIMAG(ZDUM))
c
C FIRST EXECUTABLE STATEMENT CGBFA
c
M = ML + MU + 1
INFO = 0
c
C ZERO INITIAL FILL-IN COLUMNS
c
JO = MU + 2
J1 = MINO(N,M) - 1
IF (J1 .LT. JO) GO TO 30
DO 20 JZ = JO, J1
IO = M + 1 - JZ
DO 10 I = IO, ML
zABD(I,JZ) = (O.OEO,O.OEO)
10 · CONTINUE
20 CONTINUE
30 CONTINUE
JZ = J1
JU = 0
c
c GAUSSIAN ELIMINATION WITH PARTIAL PIVOTING
c
NM1 = N - 1
IF (NM1 .LT. 1) GO TO 130
DO 120 K = 1, NM1
KP1 = K + 1
c
C ZERO NEXT FILL-IN COLUMN
c
JZ = JZ + 1
IF (JZ .GT. N) GO TO 50
IF (ML .LT. 1) GO TO 50
DO 40 I = 1, ML
zABD(I,JZ) = (O.OEO,O.OEO)
40 CONTINUE
50 CONTINUE
I-15
c
C FIND L = PIVOT INDEX
c
LM = MINO(ML N-K)
1

L = ICAMAX(LM+1 1 ZABD(M 1 K) 1 1) + M- 1
IPVT(K) = L + K - M
c
C ZERO PIVOT IMPLIES THIS COLUMN ALREADY TRIANGULARIZED
c
IF (CABS1(zABD(LIK)) .EQ. O.OEO) GO TO 100
c
C INTERCHANGE IF NECESSARY
c
IF (L .EQ. M) GO TO 60
zT = zABD (L 1 K)
zABD(L K) = zABD(M K)
1 1

zABD(M 1 K) = zT
60 CONTINUE
c
C COMPUTE MULTIPLIERS
c
zT = -(1.0E0 1 0.0EO)/zABD(M 1 K)
CALL CSCAL(LMI zTI zABD(M+1 1 K) 1 1)
c
C ROW ELIMINATION WITH COLUMN INDEXING
c
JU = MINO(MAXO(JU MU+IPVT(K)) N)
1 1

MM = M
IF (JU .LT. KP1) GO TO 90
DO 80 J = KP1 1 JU
L = L - 1
MM=MM-1
zT = zABD(L 1 J)
IF (L .EQ. MM) GO TO 70
zABD(L 1 J) = zABD(MM 1 J)
zABD(MM 1 J) = zT
70 CONTINUE
CALL CAXPY(LMI zTI zABD(M+1 1 K) 1 1 1 zABD(MM+1 1 J) 1 1)
80 CONTINUE
90 CONTINUE
GO TO 110
100 CONTINUE
INFO = K
print* 1 'L,k,abd=' ,L,k,zabd(L,k)
read(*,' (a1) ') q1
110 CONTINUE
120 CONTINUE
130 CONTINUE
IPVT(N) = N
IF (CABS1(zABD(M,N)) .EQ. O.OEO) then
INFO = N
print*,'n,zabd(m,n)= 1 n, zabd(m,n)
I

I-16
end if .; .

c
RETURN
EOO
c
c
c
SUBROUTINE CGBSL(LDA, N, ML, MU, zb)
c
C Original written on Aug. 14, 78, revised on Aug. 1, 82
C AUTHOR MOLER, C. B., (U. OF NEW MEXICO)
c
c DESCRIPTION
c
c CGBSL solves the complex band system
c A* X= B or CTRANS(A) * X = B
c using the factors computed by CGBCO or CGBFA.
c
c On Entry
c
c COMPLEX(LDA, N) , the output from CGBCo or CGBFA,
c passed by common block.
c LDA INTEGER, the leading dimension of the array ABD
c N INTEGER, the order of the original matrix.
c ML INTEGER, number of diagonals below the main diagonal.
c MU INTEGER, number of diagonals above the main diagonal.
c IPVT INTEGER(N), the pivot vector from CGBCO or CGBFA.
c passed to this subroutine by common block.
c zB COMPLEX(N)I the right hand side vector.
c
c On Return
c
c zB the solution vector X .
c
c Error Condition
c
c A division by zero will occur if the input factor contains a
c ze~o on the diagonal. Technically this indicates singularity
c but it is often caused by improper arguments or improper
c setting of LDA . It will not occur if the subroutines are
c called correctly and if CGBCO has set RCOND .GT. 0.0 or CGBFA
c has set INFO .EQ. 0 .
c
c LINPACK. This version dated 08/14/78 .
c Cleve Moler, University of New Mexico, Argonne National Lab.
c
c subroutines and Functions used in this subroutine
c
c BLAS CAXPY, zCDOTC Fortran CONJG,MINO
c
c REF: DONGARRA J.J., BUNCH J.R., MOLER C.B., STEWART G.W.,
c LINPACK USERS GUIDE, SIAM, 1979.

I-17
c
c Modified by Jerome Maa on Nov. 1996
c 1. the matrix ZABD, IPVT are transferred by common block
c 2. delete the Job parameter, so this subroutin~ only for
c solving AX=B
c 3. use implicit for complex variables
c
parameter (IW=310, kq=400)
implicit complex (z)
common /works/ zabd(iw,kq), ipvt(kq)
dimension zb(1)
c
M = MU + ML + 1
NM1 = N - 1
c
C SOLVE A * X = B, but first to do the same factoralization and
c partial pivoting on the column matrix ZB.
c
IF (ML .EQ. 0) GO TO 30
IF (NM1 .LT. 1) GO TO 30
DO K = 1, NM1
LM = MINO(ML,N-K)
L = IPVT (K)
zT = zB (L)
IF (L .EQ. K) GO TO 10
zB (L) = zB (K)
zB(K) = zT
10 CONTINUE
CALL CAXPY(LM, zT, zABD(M+1 1 K) 1 1 1 zB(K+1), 1)
end. do
30 CONTINUE
c
C . NOW SOLVE A*X = B
c
DO KB = 1 1 N
K=N+1-KB
zB(K) = zB(K)/zABD(M,K)
LM = MINO(K~M) - 1
LA = M - LM
LB = K - LM
zT = -zB (K)
CALL CAXPY(LM 1 zT, zABD(LAIK) I 1 1 zB(LB), 1)
end do
c
C Recovery sequence of unknown
c
DO 1000 KB=1,N
K=N+1-KB
L=IPVT (K)
zT=zB(L)
IF(L. EQ. K) GOTO 1000
zB (L) =ZB (K)
I-18
zB(K)=zT
1000 continue
c
RETURN
END
c
c
SUBROUTINE CSCAL(N,CA,CX,INCX)
c
c PURPOSE Complex vector scale x = a*x
c WRITTEN on Oct. 1, 79, REVISION on Aug. 01, 82
c CATEGORY NO. D1A6
c KEYWORDS BLAS,COMPLEX,LINEAR ALGEBRA,SCALE,VECTOR
c AUTHORS LAWSON, C. L., (JPL), HANSON, R. J., (SNLA)
c KINCAID, D. R., (U. OF TEXAS), and KROGH, F. T., (JPL)
c
c DESCRIPTION
c
c B L A S Subprogram
c Description of Parameters
c
c --Input--
c N number of elements in input vector(s)
c CA complex scale factor
c ex complex vector with N elements
c INCX storage spacing between elements of ex
c
c --Output--
c CSCAL complex result (unchanged if N .LE. 0)
c
c replace complex ex by complex CA*CX.
c For I = 0 to N-1, replace CX(1+I*INCX) with ,
c CA * CX(1+I*INCX)
c REF: LAWSON C.L., HANSON R.J., KINCAID D.R., KROGH F.T.,
c "BASIC LINEAR ALGEBRA SUBPROGRAMS FOR FORTRAN USAGE, 11
c ALGORITHM NO. 539, TRANSACTIONS ON MATHEMATICAL
c SOFTWARE, VOLUME 5, NUMBER 3, SEPTEMBER 1979, 308-323
c ROUTINES CALLED (NONE)
c
COMPLEX CA,CX(1)
C FIRST EXECUTABLE STATEMENT CSCAL
IF(N .LE. 0) RETURN
NS = N*INCX
DO I = 1,NS,INCX
ex ( I ) = CA *ex ( I )
end do
RETURN
END
c
c
SUBROUTINE CAXPY(N,CA,CX,INCX,CY,INCY)
c
I-19
C WRITTEN on Oct. 01, 79, REVISION on April 25, 84
C CATEGORY NO. D1A7
C KEYWORDS BLAS,COMPLEX,LINEAR ALGEBRA,TRIAD,VECTOR
c AUTHOR LAWSON c. L.
I JPL) ·, HANSON R. J.
I ( I SNLA) I (

C KINCAID, D. R., (U. OF TEXAS), KROGH, F. T., (JPL)


C PURPOSE Complex computation y = a*x + y
C DESCRIPTION
c
C B L A S Subprogram
C Description of Parameters
c
C --Input--
C N number of elements in input vector(s)
C CA complex scalar multiplier
c ex complex vector with N elements
c INCX storage spacing between elements of ex
C CY complex vector with N elements
C INCY storage spacing between elements of CY
c
C --Output--
C CY complex result (unchanged if N .LE. 0)
c
C Overwrite complex CY with complex CA*CX + CY.
C For I = 0 to N-1, replace
C CY(LY+I*INCY) with CA*CX(LX+I*INCX) + CY(LY+I*INCY),
C where
C LX= 1 if INCX .GE. 0, else LX= (-INCX)*N
C and LY is defined in a similar way using INCY.
C REFERENCES
c LAWSON c. L. HANSON R. J. KINCAID D. R. KROGH F. T.
I I I I

C *BASIC LINEAR ALGEBRA SUBPROGRAMS FOR FORTRAN USAGE*,


C ALGORITHM NO. 539, TRANSACTIONS ON MATHEMATICAL
C SOFTWARE, VOLUME 5, NUMBER 3, SEPTEMBER 1979, 308-323
C ROUTINES CALLED (NONE)
c
COMPLEX CX(1) ,CY(1) ,CA
C FIRST EXECUTABLE STATEMENT CAXPY
CANORM = ABS (REAL ( CA) ) + ABS (AIMAG ( CA) )
IF(N.LE.O.OR.CANORM.EQ.O.EO) RETURN
IF(INCX.EQ.INCY.AND.INCX.GT.O) GO TO 20
KX = 1
KY = 1
IF(INCX.LT.O) KX = 1+(1-N)*INCX
IF(INCY.LT.O) KY = 1+(1-N)*INCY
DO 10 I = 1,N
CY(KY) = CY(KY) + CA*CX(KX)
KX = KX + INCX
KY = KY + INCY
10 CONTINUE
RETURN
20 CONTINUE
NS = N*INCX
I-20
DO 30 I=1,NS,INCX
CY(I) = CA*CX(I) + CY(I)
30 CONTINUE
RETURN
END
c
c
INTEGER FUNCTION ICAMAX(N,CX,INCX)
c
C WRITTEN on Oct. 01, 79, REVISed on Aug. 01, 82
C CATEGORY NO. D1A2
C KEYWORDS BLAS,COMPLEX,LINEAR ALGEBRA,MAXIMUM COMPONENT,VECTOR
C AUTHOR LAWSON, C. L., (JPL), HANSON, R. J., (SNLA)
C KINCAID, D. R., (U. OF TEXAS), KROGH, F. T., (JPL)
c PURPOSE Find the location (or index) of the largest component
c of a complex vector
C DESCRIPTION
c
c B L A S Subprogram
c Description of Parameters
c
C --Input--
C N number of elements in input vector(s)
c ex complex vector with N elements
c INCX storage spacing between elements of ex
c
c --Output--
c ICAMAX smallest index (zero if N .LE. 0)
c
C Returns the index of the component of CX having the
c largest sum of magnitudes of real and imaginary parts.
c ICAMAX = first I, I = 1 to N, to minimize
C ABS(REAL(CX(1-INCX+I*INCX))) + ABS(IMAG(CX(1-INCX+I*INCX)))
C REF: LAWSON C.L., HANSON R.J., KINCAID D.R., KROGH F.T.,
C *BASIC LINEAR ALGEBRA SUBPROGRAMS FOR FORTRAN USAGE*,
C ALGORITHM NO. 539, TRANSACTIONS ON MATHEMATICAL
C SOFTWARE, VOLUME 5, NUMBER 3, SEPTEMBER 1979, 308-323
C ROUTINES CALLED (NONE)
c
c
COMPLEX CX(1)
C***FIRST EXECUTABLE STATEMENT ICAMAX
ICAMAX = 0
IF(N.LE.O) RETURN
ICAMAX = 1
IF(N .LE. 1) RETURN
NS = N*INCX
II = 1
SUMMAX = ABS(REAL(CX(1))) + ABS(AIMAG(CX(1)))
DO 20 I=1,NS,INCX
SUMRI = ABS(REAL(CX(I))) + ABS(AIMAG(CX(I)))
IF(SUMMAX.GE.SUMRI) GO TO 10

I-21
SUMMAX = SUMRI
ICAMAX = II
10 II = II + 1
20 CONTINUE
RETURN
END
c
c
COMPLEX FUNCTION zCDOTC(N,CX,INCX,CY,INCY)
C***BEGIN PROLOGUE zCDOTC
C***DATE WRITTEN 791001 (YYMMDD)
C***REVISION DATE 820801 (YYMMDD)
C***CATEGORY NO. D1A4
C***KEYWORDS BLAS,COMPLEX,INNER PRODUCT,LINEAR ALGEBRA,VECTOR
C***AUTHOR LAWSON, C. L., (JPL)
C HANSON, R. J., (SNLA)
C KINCAID, D. R., (U. OF TEXAS)
C KROGH, F. T., (JPL)
C***PURPOSE Dot product of complex vectors, uses complx
c conjugate of first vector
C***DESCRIPTION
c
c B L A S Subprogram
c Description of Parameters
c
c --Input--
c N number of elements in input vector(s)
c ex complex vector with N elements
c INCX storage spacing between elements of ex
c CY complex vector with N elements
c INCY . storage spacing between elements of CY
c
c --Output--
c zCDOTC complex result (zero if N .LE. 0)
c
c Returns the dot product for complex ex and CY, uses
C CONJUGATE(CX)
C zCDOTC=SUM for I=O to N-1 of CONJ(CX(LX+I*INCX))*CY(LY+I*INCY)
C where LX= 1 if INCX .GE. 0, else LX= (-INCX)*N, and LY is
C defined in a similar way using INCY.
C***REFERENCES
C LAWSON C.L., HANSON R.J., KINCAID D.R., KROGH F.T.,
C *BASIC LINEAR ALGEBRA SUBPROGRAMS FOR FORTRAN USAGE*,
C ALGORITHM NO. 539, TRANSACTIONS ON MATHEMATICAL
C SOFTWARE, VOLUME 5, NUMBER 3, SEPTEMBER 1979, 308-323
C***ROUTINES CALLED (NONE)
C***END PROLOGUE zCDOTC
c
COMPLEX CX(1) ,CY(1)
C***FIRST EXECUTABLE STATEMENT zCDOTC
zCDOTC = (0.,0.)
IF(N .LE. O)RETURN

I-22
IF(INCX.EQ.INCY .AND. INCX.GT.O) GO TO 20
KX = 1
KY = 1
IF(INCX.LT.O) KX = 1+(1-N)*INCX
IF(INCY.LT.O) KY = 1+(1-N)*INCY
DO 10 I = 1,N
zCDOTC = zCDOTC + CONJG(CX(KX))*CY(KY)
KX = KX + INCX
KY = KY + INCY
10 CONTINUE
RETURN
20 CONTINUE
NS = N*INCX
DO 30 I=1,NS,INCX
zCDOTC = CONJG(CX(I))*CY(I) + zCDOTC
30 CONTINUE
RETURN
END

I-23
Appendix II. Source List of FORTRAN Porgram P_THRIFT.FOR

program p_thrify
c
c This program solves the Laplace equation (p) in the area
c 0.0 <=X<= 3.1416, 0.0 <= y <= 3.1416.
c using the thrifty band matrix solver
c This method was first developed by Li, c. (1995), and later,
c improved and docummented by Jerome P.-Y. Maa
c
c The Govern Equation is dA2(p)/dxA2 + dA2(p)/dyA2 = o
c The finite difference equation is:
c p(i-1,j) +r*p(i,j-1) -2*(1+r)*p(i,j) +r*p(i,j+1) +p(i+1,j) o
c
c The boundary conditions are:
c p = 2 at X = 0
c p = 1 at y = 0
c p = 0 at X = y pi
c
c INFILE and OUTFILE : to store input and main output data.
c RECFILE is used to store routine checking information. If the
c program run O.K., it can be deleted.
c A MATLAB program PHIPLT.M has been developed to plot the
c output results. It uses Matlab version 4.01 for windows, and
c read output data from the OUTFILE directly.
c
c Data in the file INFILE are generated by another program
c ID.FOR for the examples given in this program.
c
c In this example, there is no need to use complex variable.
c But, it was used anyway for general applications.
c
c Jerome P.-Y. Maa Virginia Institute of Marine Science.
c
parameter (iq=110, jq=110, iw=310, jw=560, kq=10000, Lq=3)
c
implicit complex (z)
character*45 title, label
character*20 infile, outfile, recfile
character*1 id, q1
c
common /matrx/ za(kq,5), ia(kq,5), zb(kq)
common /gener/ mp, np, dx, dy, r, mu, mL, m, Lda
common /iando/ p(iq,jq), id(iq,jq)
common /unknw/ n, imap(kq), jmap(kq), irow(iq)
common /labbl/ ind(iq), jin(iq,Lq), jout(iq,Lq)
common /works/ zw(iw,jw)I ipvt(kq)
c
c boundary conditions
c
bcx0=2

II-1
bcxn=O ....
bcy0=1
bcyn=O
c
print*,'There are three resolutions for this case:'
print*,' coarse, middle, and fine resolution'
print*,'Select 1. p s.grd; 2. p_m.grd; 3. p_L.grd
read(*,*) iption -
go to (10, 12, 14), iption
10 infile='p s.grd'
outfile='p ts.out'
recfile='p-ts.rec'
go to 20
12 infile='p m.grd'
outfile='p_tm.out'
recfile='p_tm.rec'
go to 20
14 infile='p L.grd'
outfile='p tL.out'
recfile='p=tL.rec'
20 continue
c
open(9, file=infile, status='old')
open(10, file=outfile, status='unknown')
open(11, file=recfile, status='unknown')
c
c mp,np Max. grid numbers in x and y direction, respectively.
c dx,dy grid sizes in x and y direction, respectively.
c
read(9,'(a45)') title
write(11,' (a45) ') title
read(9,*) mp,np,dx,dy
if(mp .gt. iq) then
print*,'IQ is smaller than MP, Change IQ to' ,mp
stop
end if
if(np .gt. jq) then
print*,'JQ is smaller than NP, Change JQ to' ,np
stop
end if
c
r = (dx/dy)**2
c
c ID code for each grid point
c 0, interior point
c
c 1, upper boundary condition
c 2, lower boundary condition
c 3, left boundary condition
c 4, right boundary condition
c
c 5, left bottom corner B.C.

II-2
c 6, left top corner B.C.
c 7, right bottom corner B.C.
c 8, right top corner B.C.
c
c e, grid point that is not included in the study domain
c
c The following is an example
c
c j
A
c np 6111111111118eeeeeeeeeeeeeeeeeee
c 3000000000004eeeeeeeeeeeeeeeeeee
c 30000000000001111111111111111118
c 30000000000000000000000000000004
c 30000000000000000000000000000004
c 30000000000000000000000000000004
c 30000000000000000000000000000004
c 52200000000000000000000000222227
c eee30000000000000000000004eeeeee
c eee30000000000000000000004eeeeee
c eee30000000000000000000004eeeeee
c 1 eee52222222222222222222227eeeeee
c 1 mp
c ----------------------------------------------> i
c
read(9,30) label
write(11,30) label
read(9,30) label
write(11,30) label
30 format (a50)
do j=1,np
jj=np-j+1
read(9,45) jns, (id(i,jj) ,i=1,mp)
45 format(iS, (110a1) )
if(jns .ne. jj) then
write(11,50) jj,jns
50 format(' Sequence is wrong for ID input at' ,2i5)
stop
end if
end do
print*, 'Completed reading ID code matrix'
close(9)
c
c construct the unknow COLUMN matrix X, in the full matrix eq.
c AX=B. Find each unknown's location: imap(map) ,jmap(map)
c
c ind(i) : no. of isolated sector in each culornn, x grid.
c for this particular case, ind(i) are all 1 because
c no land points in the middle of study domain.
c jin(i,index): begin grid number for an isolated sector in a
c column
c jout(i,index): end grid number for an isolated sector in a
c culornn

II-3
c map the number of total unknow, or the length of X.
c later, it is reassigned as N
c irow (i) the total number of unknow, P, in each column
c
map=O
do i=1,mp
icount=O
in=O
iout=1
index=1
do j=l,np
if ( id ( i, j) . eq. '0' ) then
icount=icount+1
map=map+1
imap(map)=i
jmap(map)=j
if(in .eq. 0) then
jin(i,index)=j
in=1
iout=O
end if
end if
c
if( id(i,j) .eq. '1' .and. iout .eq. 0) then
jout(i,index)=j
iout=l
in=O
index=index+l
end if
end· do
irow(i)=icount
ind(i)=index-1
c
c ind(i) should be >= 1, except for the entire column are all
c boundary points. if not, something wrong.
c
write(11,55) i, ind(i), irow(i), jin(i,l), jout(i,l)
55 format(' i,ind,irow,jin,jout=' ,SiS)

if( ind(i) .gt. Lq) then


write(*,60) i, ind(i), Lq, ind(i)
60 format(' At i=' ,i4,' Ind(i)=' ,i2,' > Lq=' ,i2/
* ' Change Lq to' ,i3,' and re-run')
stop
end if
end do
c
c n: length of the banded matrix
c
n=map
if(n .lt. O.S*kq) then
write(*,65) n, kq

II-4
6S format('----------------------------------- -----------
'/
N=' ,iS,' << KQ=' ,iS//' It is better to reduce KQ'
* 1

* ' length of the Banded Matrix,'/' So that KQ is not''


* ' >>than that required.'/' Instead, one should' '
* ' increase the length of the working matrix, JW,:/
* ' in order to reduce disk I/O and computing time.'/
* ' You may continue (c), or re-run (r), < c/r > : ')
read ( * , ' ( a1) ' ) q1
if( q1 .eq. 'r' .or. q1 .eq. 'R') stop
end if
if(n .gt. kq) then
write(*,70) n, kq, n
70 format('----------------------------------------------' I
* 'N=' ,iS,' > KQ=' ,iS//
* ' Increases KQ t o ' ,iS, ' and re-run')
stop
end if
c
c Set up the two small matrices ZA and IA for storing the banded
c matrix and find out the band width, work size, etc.
c
mu=O
mL=O
write ( 11, 8 0 )
80 format (' i j ia ( 1) , .... , ia ( 4) , ia ( s)
* ' za ( 1) . . . za ( S) zb')
do map=1,n
i=imap(map)
j =j map (map)
c
call domain(i, j, map, ierr)
if( ierr .eq. 0) then
write(11,90)i,j,ia(map,1) ,ia(map,2) ,ia(map,3) ,ia(map,4),
* ia(map,S) 1 real(za(mapl1)) 1 real(za(mapl2)) I
* real(za(map,3)), real(za(mapl4)) I real(za(map 1 S))
1

* real(zb(map) )
90 format(1x,2iS,2x,SiS,2x 1 6f6.2)
else
print* 1 'stop 1 error in DOMAIN at iljlmap=' I i,j 1 map
stop
end if
c
c find the band width
c
if( ia(map 1 S) .ne. 0 ) then
mu c=abs(ia(mapiS) - ia(map,3)
else -
mu c=O
end if
if(mu .lt. mu c) mu=mu c
if( ia(map 1 1) .ne. 0 ) then
mL_c=abs(ia(mapl1} - ia(mapl3}

I I-S
else
mL c=O
end if
if(mL .lt. mL_c) mL=mL c
end do
c
m=mL+mu+1
Lda=m+mL
write(*,100) m, n, mu, mL, Lda
100 format(' Band width, m = ',i7/' Data point, N = ',i7/
* ' Upper B.W., mu = ',i7/' Lower B.W., mL = ',i7/
* ' Lda for Zw = ',i7)
if(Lda .gt. iw) then
print*,'Please increase IW to ',Lda
stop
end if
c
c uses the complex thrifty band matrix solver to solve ZAF*ZX= ZB
c
call bmsolver(ier_code)
c
if(ier code .eq. 0) then
do map=1, n
i=imap(map)
j=jmap(map)
p(i,j)=real(zb(map)
end do
do j=1,np
p(1,j) = bcxO
p(mp,j) = bcxn
end do
do i=1,mp
p(i,1) = bcyO
p(i 1 np) = bcyn
end do

p(1 1 1) = O.S*(bcyO+bcxO)
p(1,np) = O.S*(bcyn+bcxO)
p(mp,1) = O.S*(bcyO+bcxn)
p(mp 1 np) = O.S*(bcyn+bcxn)

write (10,' (a45) title


1
)

wrfte(10, 1 (2i5,2f12.8) ') mp 1 np 1 dx 1 dy

do j=1,np
write ( 10 I 12 0 ) j , (p ( i , j ) 1 i =1, mp)
120 format(i5/(10f8.4))
end do
end if

close(10)
close(11)

II-6
print*, 'program stop'
print*, 'The solution is in the file ',outfile
print*,'and oher details, if needed, is in recfile
stop
end
c
c
*****************************************************************
*
c
subroutine BMsolver(ierr code)
c
c It solves a complex banded matrix equation ZAF * ZX = ZB
c where ZAF is a complex band matrix with dimension (m * n)
c ZX and ZB are two complex column matrices with length (n) .
c
c Because of the hudge size of ZAF, e.g., (300 x 50000), it is
c designed to solve this problem using the following two steps.
c
c First, don't use the full size of ZAF, instead, uses two small
c matrices: ZA & IA that each only uses 50000 x 5 to save space.
c
c ZA : a complex matrix (n * 5) to store the coefficent matrix in
c a matrix equation ZAF*ZX= ZB. Because of using the finite
c difference method to solve an elliptic equation, there are
c only 5 elements to be saved for the coefficient matrix.
c The band width, however, is much much large than 5 with a
c lot of "zero." By doing so, we need another matrix
c IA : to save the corresponding locations.
c
c Second, uses a working matrix, ZW(IW,JW), and do a systematical
c
c swap between a hard disk and memory. For this reason, be sure
c that you do have enough space in your hard disk. For example, a
c complex matrix with size of (300 x 50000) requires 120 MB for
c storage, if using 4 byte for a real number. If using 8 bytes,
c then, 240 MB is needed.
c
c IW,JW: IW should be >= m+ml, where ml is the lower band width.
c The size of JW depends on the available computer memory.
c In general, the large the JW, the less the disk IO, and
c thus, the faster the computing speed. As a rule of
c thumb, you may select JW = 2*IW and tried to see if your
c computer has enough memory to run the program.
c
c The proceedures follows that given in the subroutine CGBFA &
c CGBSL from LINPACK. The major difference is just doing it one
c block at a time, stores the results in hard disk sequently.
c After forward elimination, reverse the process by reading the
c data from hard disk and do back substitute for the solution.
c
parameter (iq=llO, jq=llO, iw=310, jw=560, kq=lOOOO)

II-7
c
implicit complex (z)
character*8 tmpfile
character*3 chrc
character*l id, ql
c
common /matrx/ za(kq,S), ia(kq,S), zb(kq)
common /gener/ mp, np, dx, dy, r, mu, mL, m, Lda
common /iando/ p(iq,jq), id(iq,jq)
common /unknw/ n, imap(kq), jmap(kq), irow(iq)
common /works/ zw(iw,jw), ipvt(kq)
c
cabsl(zdum)=abs(real(zdum)) + abs(aimag(zdum))
c
c Gaussian elimination with partial pivoting
c
c NK: # of columns for the working matrix that will be saved.
c NW: The number of column needed for working, nw=m-l+nk
c NS: An index to show the number of working matrix.
c
c check working matrix dimension
c
nk=3SO
nw=m+nk-1
c
if(nw .gt. jw) then
write(*,lO) iw, jw, Lda, nw
:o format(' The working matrix, ZW(iw,jw), iw,jw=', 2iS/
* ' IW shoult be >=',iS, ' and JW must be >=',iS/
* ' Please change program and re-run .'//
* ' If no memory, you can reduce the NK')
ierr code=l
return
end if
c
c clear up the working matrix
c
do i=l,iw
do j=l,nw
ZW ( i 1 j ) = ( 0 • 0 1 0 • 0 )
end do
end do
c
c construct the first working matrix, ZW, which has a size of
c (Lda x nw) . The thrify storaged matrices are expanded, only
c to the first (nk x S) block.
c
c The flag is used to recording only once when the ZA data
c starts lost.
c
ns=O
mapl=O

II-8
iflag=1
ne=nw
if(ne .gt. n) ne=n
do map=1,ne
do i=1,5
j=ia(map,i)
if(j .ne. 0) then
if(j .le. nk+m-1) then
ids=j-map
zw(m-ids,j)=za(map,i)
else
if(iflag .eq. 1) then
map1=map
iflag=O
end if
end if
end if
end do
end do
c
c Gaussian elimination with partial pivoting
c
j1=minO(n,m)-1
jz=j1
ju=O
nt=O
index=O
c
100 continue
c
if(ju .ne. 0) ju=ju-nk
if(jz .ne. j1) jz=jz-nk
c
c index=O, the first, 2nd, ... , block of matrix.
c =1, the last block of matrix
c
if(index .eq. 0 .and. ne .ne. n) then
nc=nk
else
nc=n-ns*nk-1
end if
c
do k=1,nc
kp1=k+1
kr=ns*nk + k
c
c find L = pivot index
c
Lm=minO(mL,n-ns*nk-k)
L=icamax(Lm+1, zw(m,k), 1) + m -1
ipvt(kr)=L+kr-m
c

II-9
c zero pivot implies that this column are all zeros, a singular
c matrix.
c
if (cabs1 (zw (L, k)) .lt. 0 .10e-8) goto 120
c
c interchange if necessary
c
if(L .ne. m) then
zt==ZW(L,k)
zw(L,k)==zw(m,k)
zw(m,k)==zt
end if
c
c compute multipliers
c
zt==-(1.0eO,O.OeO)/zw(m 1k)
call cscal(Lm 1 zt 1 zw(m+11k) I 1)
c
c swap ZB array, if necessary
c
Lp==ipvt (kr)
zt==zb(Lp)
if(Lp .ne. kr) then
zb(Lp)==zb(kr)
zb(kr)==zt
end if
call caxpy(Lml. ztl zw(m+11k) I 11 zb(kr+1), 1)
c
c row elimenation with column indexing
c
ju=minO(maxO(ju, mu+ipvt(kr)-ns*nk), n-ns*nk)
mm=m
if(ju .ge. kp1) then
do j=kp1,ju
L=L-1
mm=mm-1
zt=zw(L,j)
if(L .ne. mm) then
zw(L,j)=zw(mm,j)
zw(mm,j)=zt
end if
call caxpy(Lml ztl zw(m+1,k), 1, zw(mm+1,j), 1)
end do
end if
c
goto 150
120 continue
print*,'Zero diagonal element at (L,k)=', L, k
stop
150 continue
c
end do

II-10
c
c Eexcept the last working matrix, write the upper triangular
c matrix (from j=1,nk) into hard disk. For the last one, i.e.,
c nc<>nk, go to back substitute directly.
c
if(nc .eq. nk) then
nt=nt+1
print*,'Writing tern. file#', nt
tmpfile='t'//chrc(nt)//' .tmp'
open(12,file=tmpfile,form='unformatted')
write ( 12 ) ( ( z w ( i , j ) , i = 1 , m) , j =1 , nk)
close(12)
c
c moving the rest working matrix forward to the beginning
c
do j=nk+1, nk+m-1
L=j-nk
do i=1,Lda
zw(i,L)=zw(i,j)
end do
end do
c
c clear the rest working area for reading new working matrix
c
do j=m, nk+m-1
do i=1,Lda
ZW ( i j ) = (0 • 0 0 • 0 )
1 1

end do
end do
c
c read in a full block of ZA and IA, by two steps
c
ns=ns+1
if( (ns+1)*nk+m-1 .lt. n) then
ne=nk
index=O
c
c For intermediate blocks, read in ZA and IA by two steps.
c First read in the upper triangular matrix that were cut off
c at the previous time when constructing the working matrix.
c
if(map1 .ne. 0) then
do map=map1, m-1+ns*nk
do i=4,5
j=ia(map,i)
if(j .ne. 0) then
if(j .gt. m-1+ns*nk) then
ids=j-map
zw(m-ids,j-ns*nk)=za(map,i)
end if
end if
end do

II-11
end do
end if
c
c now read in next block of ZA and IA
c
iflag=1
map1=0
do k=1,ne
map=k + ns*nk + m-1
do i=1,5
j=ia(map,i)
if(j .ne. 0) then
if(j .le. ns*nk+ne+m-1) then
ids=j-map
zw(m-ids,j-ns*nk)=za(map,i)
else
if(iflag .eq. 1) then
map1=map
iflag=O
end if
end if
end if
end do
end do
c
c End of read in a block of ZA and IA.
c
else
c
c This is to read the last block of ZA and IA
c
ne=n-(m-1+ns*nk)
index=1
if(map1 .ne. 0) then
do map=map1,m-1+ns*nk
do i=4,5
j=ia(map,i)
if(j .ne. 0) then
if(j .gt. m-1+ns*nk) then
ids=j-map
zw(m-ids,j-ns*nk)=za(map,i)
end if
end if
end do
end do
end if
c
c now reads the last block of ZA and IA
c
iflag=1
map1=0
do k=1,ne

II-12
map=m-1+ns*nk+k
do i=1,5
j=ia(map,i)
if(j .ne. 0) then
if(j .le. m-1+ns*nk+ne) then
ids=j-map
zw(m-ids,j-ns*nk)=za(map,i)
else
ierr code=2
print*,' If this happen, it is wrong.' .
print*, 'At the last block, no overflow'
return
end if
end do
end do
iflag=1
end if
c
goto 100
c
else
c
c backward substitution from the last submatrix
c
if( nt .eq. 0) nc=n-1
do kb=1,nc+1
kr=n+1-kb
k=nc+2-kb
zb(kr)=zb(kr)/zw(m,k)
Lm=minO(kr,m)-1
La=m-Lm
Lb=kr-Lm
zt=-zb(kr)
call caxpy(Lm, zt, zw(La,k), 1, zb(Lb), 1)
end do
c
c If one loop can include all elements, i.e., for a small banded
c matrix, just stops after this.
c
if( nt .eq. 0 ) go to 400
end if
c
c complete the rest backward substitute
c
ns=O
200 continue
c
c clear the working matrix for reading new submatrix from disk.
c
do j=1,nk+m-1
do i=1,m
zw ( i , j ) = ( 0 . 0 , 0 . 0 )

II-13
end do
end do
c
print*1 1Reading tern. file # 1
1 nt
open(121file=tmpfile 1 form= 1unformatted 1 )
read(12) ((zw(i 1 j) 1 i=1,m) ,j=1 1nk)
close(12 1 status='delete')
c
do kb=1 1nk
kr=n+1-(nc+1)-ns*nk-kb
k=nk+1-kb
zb(kr)=zb(kr)/zw(m,k)
Lm=minO(kr,m)-1
La=m-Lm
Lb=kr-Lm
zt=-zb(kr)
call caxpy(Lm, zt, zw(La,k) I 1, zb(Lb), 1)
end do
ns=ns+1
nt=nt-1
tmpfile='t'//chrc(nt)//' .tmp'
if(nt .gt. 0) goto 200

400 continue
c
c To restore the original sequence of ZB. Since it starts
c at the 2nd. we started at 2nd too for restoring.
c
print*,'Restore the original sequency'
do 1000 kb=2 1n
k=n+1-kb
L=ipvt (k)
zt=zb (L)
if( L .eq. k ) go to 1000
print*,'pivoting at L,k,kb= 1 ,L,k 1kb
zb(L)=zb(k)
zb(k)=zt
1000 continue
c
return
end
c
c
character*3 function chrc(nt)
c
c It changes an input integer number NT to character
c CHANGQING LI 06/94
c
character*1 c1
character*2 c2
character*3 c3
c

II-14
i1==nt
i2==i1/10
if(i2 .gt. 0) then
i3==i2/10
if(i3 .gt. 0) then
i4==i3/10
if(i4 .gt. 0) then
print*,'----------------------------------------'
print*,'The given integer is> 999, not allowed.'
print*,'---------------------------------------- 1

chrc=='-1'
else
c3==char(48+i3)//char(48+i2-i3*10)//char(48+i1-i2*10)
chrc==c3
end if
else
c2==char(48+i2)//char(48+i1-i2*10)
chrc==c2
end if
else
c1==char(48+i1)
chrc==c1
end if
c
return
end
c
c
SUBROUTINE CSCAL(N,CA,CX,INCX)
c
c PURPOSE Complex vector scale x == a*x
C WRITTEN on Oct. 1, 79, REVISION on Aug. 01, 82
C CATEGORY NO. D1A6
C KEYWORDS BLAS,COMPLEX,LINEAR ALGEBRA,SCALE,VECTOR
C AUTHORS LAWSON, C. L., (JPL), HANSON, R. J., (SNLA)
C KINCAID, D. R., (U. OF TEXAS), and KROGH, F. T., (JPL)
c
C DESCRIPTION
c
c B L A S Subprogram
c Description of Parameters
c
c --Input--
C N number of elements in input vector(s)
c CA complex scale factor
c ex complex vector with N elements
c INCX storage spacing between elements of ex
c
c --Output--
C CSCAL complex result (unchanged if N .LE. O)
c
c replace complex ex by complex CA*CX.

II-15
C For I = 0 to N-1, replace CX(1+I*INCX) with
C CA * CX (l+I*INCX)
C REF: LAWSON C.L., HANSON R.J., KINCAID D.R., KROGH F.T.,
C "BASIC LINEAR ALGEBRA SUBPROGRAMS FOR FORTRAN USAGE,"
C ALGORITHM NO. 539, TRANSACTIONS ON MATHEMATICAL
C SOFTWARE, VOLUME 5, NUMBER 3, SEPTEMBER 1979, 308-323
C ROUTINES CALLED (NONE)
c
COMPLEX CA,CX(1)
C FIRST EXECUTABLE STATEMENT CSCAL
IF(N .LE. 0) RETURN
NS = N*INCX
DO I = 1,NS,INCX
CX(I) = CA*CX(I)
end do
RETURN
END
c
c
SUBROUTINE CAXPY(N,CA,CX,INCX,CY,INCY)
c
C WRITTEN on Oct. 01, 79, REVISION on April 25, 84
C CATEGORY NO. D1A7
C KEYWORDS BLAS,COMPLEX,LINEAR ALGEBRA,TRIAD,VECTOR
C AUTHOR LAWSON, C. L., (JPL), HANSON, R. J., (SNLA)
C KINCAID, D. R., (U. OF TEXAS), KROGH, F. T., (JPL)
C PURPOSE Complex computation y = a*x + y
C DESCRIPTION
c
C B L A S Subprogram
C Description of Parameters
c
C --Input--
C N number of elements in input vector(s)
C CA complex scalar multiplier
C ex complex vector with N elements
C INCX storage spacing between elements of ex
C CY complex vector with N elements
C INCY storage spacing between elements of CY
c
C --Output--
C CY complex result (unchanged if N .LE. 0)
c
C Overwrite complex CY with complex CA*CX + CY.
C For I = o to N-1, replace
c CY(LY+I*INCY) with CA*CX(LX+I*INCX) + CY(LY+I*INCY),
c where
c LX= 1 if INCX .GE. 0, else LX= (-INCX)*N
C and LY is defined in a similar way using INCY.
C REFERENCES
c LAWSON c. L. HANSON R. J. KINCAID D. R. KROGH F. T.
I I I I

C *BASIC LINEAR ALGEBRA SUBPROGRAMS FOR FORTRAN USAGE*,


II-16
c ALGORITHM NO. 539, TRANSACTIONS ON MATHEMATICAL
c SOFTWARE, VOLUME 5, NUMBER 3, SEPTEMBER 1979 308-323
c ROUTINES CALLED (NONE) I

c
COMPLEX CX(1) ,CY(1) ,CA
c FIRST EXECUTABLE STATEMENT CAXPY
CANORM = ABS(REAL(CA)) + ABS(AIMAG(CA))
IF(N.LE.O.OR.CANORM.EQ.O.EO) RETURN
IF(INCX.EQ.INCY.AND.INCX.GT.O) GO TO 20
KX = 1
KY = 1
IF(INCX.LT.O) KX = 1+(1-N)*INCX
IF(INCY.LT.O) KY = 1+{1-N)*INCY
DO 10 I = 1,N
CY(KY) = CY(KY) + CA*CX(KX)
KX = KX + INCX
KY = KY + INCY
10 CONTINUE
RETURN
20 CONTINUE
NS = N*INCX
DO 30 I=1,NS,INCX
CY(I) = CA*CX(I) + CY(I)
30 CONTINUE
RETURN
END
c
c
INTEGER FUNCTION ICAMAX(N,CX,INCX)
c
C WRITTEN on Oct. 01, 79, REVISed on Aug. 01, 82
C CATEGORY NO. D1A2
C KEYWORDS BLAS,COMPLEX,LINEAR ALGEBRA,MAXIMUM COMPONENT,VECTOR
C AUTHOR LAWSON, C. L., (JPL), HANSON, R. J., (SNLA)
C KINCAID, D. R., (U. OF TEXAS), KROGH, F. T., (JPL)
c PURPOSE Find the location (or index) of the largest component
c of a complex vector
C DESCRIPTION
c
c B L A S Subprogram
c Description of Parameters
c
C --Input--
C N number of elements in input vector(s)
c ex complex vector with N elements
c INCX storage spacing between elements of ex
c
c --Output--
c ICAMAX smallest index (zero if N .LE. 0)
c
C Returns the index of the component of CX having the
c largest sum of magnitudes of real and imaginary parts.

II-17
C ICAMAX = first I, I = 1 to N, to m~n~m~ze
C ABS(REAL(CX(1-INCX+I*INCX))) + ABS(IMAG(CX(1-INCX+I*INCX)})
C REFERENCES
C LAWSON C.L., HANSON R.J., KINCAID D.R., KROGH F.T.,
C *BASIC LINEAR ALGEBRA SUBPROGRAMS FOR FORTRAN USAGE*,
C ALGORITHM NO. 539, TRANSACTIONS ON MATHEMATICAL
C SOFTWARE, VOLUME 5, NUMBER 3, SEPTEMBER 1979, 308-323
C ROUTINES CALLED (NONE)
c
c
COMPLEX CX(1)
C***FIRST EXECUTABLE STATEMENT ICAMAX
ICAMAX = 0
IF(N.LE.O) RETURN
ICAMAX = 1
IF(N .LE. 1) RETURN
NS = N*INCX
II = 1
SUMMAX = ABS(REAL(CX(1))) + ABS(AIMAG(CX(1)))
DO 20 I=1,NS,INCX
SUMRI = ABS(REAL(CX(I))) + ABS(AIMAG(CX(I}))
IF(SUMMAX.GE.SUMRI) GO TO 10
SUMMAX = SUMRI
ICAMAX = II
10 II = II + 1
20 CONTINUE
RETURN
END
c
c--------------------------------------------------------
c
subroutine domain(i, j, map, ier)
c
parameter (iq=110, jq=110, kq=10000, Lq=3)
implicit complex (z)
character*1 id
common /matrx/ za(kq,5), ia(kq,5), zb(kq)
common /gener/ mp, np, dx, dy, r, mu, mL, m, Lda
common /iando/ p(iq,jq), id(iq,jq)
common /unknw/ n,imap(kq) ,jmap(kq) ,irow(iq)
common /labbl/ ind(iq) ,jin(iq,Lq) ,jout(iq,Lq)
c
dimension iid(9)
c
c The finite difference governing equation, the Laplace equation,
c with r = (dx/dy)**2, is
c
c p(i-1,j) +r*p(i,j-1) -(2+2r)*p(i,j) +r*p(i,j+1) +p(i+1,j) = 0
c
c In storage, the coefficient of p(i-1,j) is stored in (map,1)'
c the coefficient of p(i,j-1) is stored in (map,2)'
c the coefficient of p(i,j) is stored in (map,3),
II-18
c the coefficient of p(i,j+1) is stored 1'n (map, 4) ,
c and the cbefficient of p(i+1,j) is stored in (map,S)
c
ier=O
do k=1,9
iid(k)=O
end do
icount=O
c
c For those grid point that are not neighbored to a boundary
c
if ( id ( i -1 j) • eq • 0
I I I and • id ( i+1 j) . eq . ' 0 ' . and .
• I

* id(i,j-1) .eq. '0' .and. id(i,j+1) . eq. ' 0 ' ) then


za(map,1)=1.0
za(map,2)=r
za(map,3)=-2.0*(1.0+r)
za(map,4)=r
za(map,5)=1.0
ia(map,1)=map-idistl(i,j)+1
ia(map,2)=map-1
ia(map,3)=map
ia(map,4)=map+1
ia(map,S)=map+idistr(i,j) -1
zb(map)=(O.O, 0.0)
icount=icount+1
iid(1)=1
end if
c
c for those has a Boundary grind point on the left
c
if(id(i-1,j) .eq. '3' .and. id(i,j-1) .eq. '0' .and.
* id ( i, j +1) . eq. '0' ) then
za(map,1)=0.0
za(map,2)=r
za(map,3)=-2.0*(1.0+r)
za(map,4)=r
za(map,5)=1.0
ia(map,1)=0
ia(map,2)=map-1
ia(map,3)=map
ia(map,4)=map+1
ia(map,S)=map+idistr(i,j) -1
zb(map)={-2.0, 0.0)
icount=icount+1
iid(2)=1
end if
c
c for those has a Boundary grind point on the right
c
if(id(i+1,j) .eq. '4' . and. id ( i, j -1) . eq. '0' .and .
* id ( i, j +1) . eq. '0' ) then
za(map,1)=1.0

II-19
za(map,2)=r
za(map,3)=-2.0*(1.0+r)
za(map,4)=r
za(map,S)=O.O
ia(map,1)=map-idistl(i,j)+1
ia(map,2)=map-1
ia(map,3)=map
ia(map,4)=map+1
ia(map,S)=O
zb(map)=(O.O, 0.0)
icount=icount+1
iid(3)=1
end if
c
c for those has a Boundary grind point on the top
c
if(id(i,j+1) .eq. '1' .and. id(i-1,j) .eq. '0' .and.
* id(i+1,j) .eq. '0' ) then
za(map,1)=1.0
za(map,2)=r
za(map,3)=-2.0*(1.0+r)
za(map,4)=0
za(map,5)=1.0
ia(map,1)=map-idistl(i,j)+l
ia(map,2)=map-l
ia(map,3)=map
ia(map,4)=0
ia(map,S)=map+idistr(i,j) -1
zb(map)=(O.O, 0.0)
icount=icount+l
iid(4)=1
end if
c
c for these grid points that has a Boundary at bottom
c
if(id(i,j-1) .eq. '2' .and. id(i-l,j) .eq. '0' .and.
* id ( i+l, j) . eq. '0' ) then
za(map,1)=1.0
za(map,2)=0
za(map,3)=-2.0*(1.0+r)
za(map,4)=r
za(map,S)=l.O
ia(map,1)=map-idistl(i,j)+l
ia(map,2)=0
ia(map,3)=map
ia(map,4)=map+l
ia(map,S)=map+idistr(i,j)-1
zb(map)=(-1.0, 0.0)
icount=icount+l
iid(5)=1
end if
c

II-20
c For those grid points that are left bottom corner
c
if(id(i-1,j) .eq. '3' .and. id(i,j-1) .eq. '2') then
za(map,1)=0.0
za(map,2)=0
za(map,3)=-2.0*(1.0+r)
za(map,4)=r
za(map,5)=1.0
ia(map,1)=0
ia(map,2)=0
ia(map,3)=map
ia(map,4)=map+1
ia(map,S)=map+idistr(i,j)-1
zb(map)=(-3.0, 0.0)
icount=icount+1
iid(6)=1
end if
c
c For those grid points that are left top corner Boundary points
c
if(id(i-1,j) .eq. '3' .and. id(i,j+1) .eq. '1') then
za(map,1)=0.0
za(map,2)=r
za(map,3)=-2.0*(1.0+r)
za(map,4)=0.0
za(map,5)=1.0
ia(map,1)=0
ia(map,2)=map-1
ia(map,3)=map
ia(map,4)=0
ia(map,S)=map+idistr(i,j)-1
zb(map)=(-2.0, 0.0)
icount=icount+1
iid(7)=1
end if
c
c For those grid points that are right bottom boundary points
c
if(id(i+1,j) .eq. '4' .and. id(i,j-1) .eq. ' 2 ' ) then
za(map,1)=1.0
za(map,2)=0
za(map,3)=-2.0*(1.0+r)
za(map,4)=r
za(map,S)=O.O
ia(map,1)=map-idistl(i,j)+1
ia(map,2)=0
ia(map,3)=map
ia(map,4)=map+1
ia(map,S)=O
zb(map)=(-1.0, 0.0)
icount=icount+1
iid(8)=1

II-21
.: .
end if
c
c For those grid points that are right top boundary points
c
if ( id ( i+1, j) . eq. '4' . and. id ( i, j +1) . eq. '1' ) then
za(map,1)=1.0
za(map,2)=r
za(map,3)=-2.0*(1.0+r)
za(map,4)=0
za(map,S)=O.O
ia(map,1)=map-idistl(i,j)+1
ia(map,2)=map-1
ia(map,3)=map
ia(map,4)=0
ia(map,S)=O
zb(map)=(O.O, 0.0)
icount=icount+1
iid(9)=1
end if
c
if( icount .ne. 1) then
ier=1
write(11,10) icount, (iid(k) ,k=1,9)
10 format(' icount=' ,iS,' id(1)-id(9)=' ,9i2/
* ' Only one of the ids can be 1, check ID array')
end if
return
end
c
c----------------------------------------------------------------
c
integer function idistr(i,j)
parameter (iq=110, Lq=3)
common /labbl/ ind(iq) ,jin(iq,Lq) ,jout(iq,Lq)
c
c Calculates the distance between points (i+1,j) and (i,j). It
c means how many grid points, which are looking for, are in
c between. It counts vertically from the point(i,j) and up
c (i,j+1), (i,j+2), ... , (i,np), then (i+1,1), (i+1,2), ... , to
c (i+1,j). In the matrix equation, this distance represents the
c up band width at that particular diagonal element (i,j)
c
do k=1,ind(i)
if(j .ge. jin(i,k) .and. j .le. jout(i,k) ) mark1=k
end do
c
do k=1,ind(i+1)
if(j .ge. jin(i+1,k) .and. j .le. jout(i+1,k) ) mark2=k
end do
c
idd=jout(i,mark1)-j+1 + j-jin(i+1,mark2)
c

II-22
do k=mark1+1, ind(i)
idd=idd + jout(i,k)-jin(i,k)+1
end do
c
do k=1,mark2-l
idd=idd + jout(i+1,k)-jin(i+1,k)+l
end do
c
idistr=idd
c
return
end
c
c
integer function idistl(i,j)
c
parameter (iq=110, Lq=3)
common /labbl/ ind(iq) ,jin(iq,Lq) 1 jout(iq 1Lq)
c
c Calculate the distance between point (i-1,j) and (i,j). The
c distance means how many grid points, which are solved for, are
c in between. It counts vertically from (i,j) downward (i,j-1),
c (i 1 j-2), ... , (i 11) and restarted from previous i line
c (i-1,np) 1 (i-1,np-1)
1 ••• to (i-1,j).
,

c
do k=l,ind(i-1)
if(j .ge. jin(i-1,k) .and. j .le. jout(i-1,k)) markl=k
end do
c
do k=1, ind (i)
if(j .ge. jin(ilk) .and. j .le. jout(i,k)) mark2=k
end do
c
idd=jout(i-1 1mark1)-j+1 + j-jin(ilmark2)
c
do k=mark1+11ind(i~1)
idd=idd+jout(i-11k)-jin(i-l,k)+1
end do
c
do k=1,mark2-1
idd=idd+jout(i,k)-jin(i,k)+1
end do
c
idistl=idd
return
end

II-23
Appendix III. Source List of FORTRAN Program ID.FOR
program id matrix
c
c This program generates the input ID matrix, I.D. codes for
c each cell
c
parameter (iq=400,jq=200, kq=SOO, Lq=2000)
character*1 id
character*12 outfile
character*30 header
character*SO title
c
common/const/ mp, np, dx, dy, r
common/array/ id(iq,jq)
common/bound/ nobc, ibc(Lq), jbc(Lq), zp(Lq)
c
c mp mesh number in x direction
c np mesh number in y direction
c dx spatial step in x-direction
c dy spatial step in y-direction
c r square of dx/dy
c
c ID code for water depth
c 0, interior cell
c 1, upper boundary condition
c 2, lower boundary condition
c 3, left boundary condition
c 4, right boundary condition
c 5, left bottom corner B.C.
c 6, left top corner B.C.
c 7, right bottom corner B.C.
c 8, right top corner B.C.
c e, land point
c
title=' Laplace equation with specified Dirichlet B.C.'
print*,'Select 1. large grid; 2. middle grid; '
print*,' 3. small grid
read(*,*) iption
go to (10, 12, 14) iption
10 mp=101
np=101
outfile='p_L.grd'
go to 20
12 mp=62
np=62
outfile='p_m.grd'
go to 20
14 mp=12
np=12
outfile='p_s.grd'
20 continue

III-1
c
·dx=3 .1415926/ (mp-1)
dy=3.1415926/(np-1)
c
c velocity potential
c
if(mp .ge. iq .or. np .ge. jq) write(*,30) mp,np, iq jq
30 format(' Given array (mp,np)=' ,2i5,' is> than alloc~ted'
* (iq,jq)=',2i5) I

c
open(8,file=outfile,form='formatted')
write(8,' (a80) ') title
write(8,40) mp, np, dx, dy
40 format(2i5, 2f12.8)
c
nobc=O
c
c first establish the ID code for the entire grid point
c
do j=1, np
do i=1, mp
id ( i j ) 0
1 =I I

end do
end do
c
c given boundary at the right hand side
c
do j=2,np-1
nobc=nobc+1
id(mp,j)='4'
ibc(nobc)=mp
jbc(nobc)=j
zp(nobc)=O
end do
c
c left side B.C.
c
do j=2,np-1
id(1,j)='3'
nobc=nobc+1
ibc(nobc)=1
jbc(nobc)=j
zp(nobc)=2.0
end do
c
c top B.C.
c
do i=2,mp-1
id(i,np)='1'
nobc=nobc+1
ibc(nobc)=i
jbc(nobc)=np

III-2
zp(nobc)=O.O
end do
c
c bottom B.C.
c
do i=2,mp-1
id(i,1)='2'
nobc=nobc+1
ibc(nobc)=i
jbc(nobc)=1
zp(nobc)=1.0
end do
c
c check the array assignment
c
if(nobc .gt. Lq) then
write(*,SO) nobc, Lq, nobc
SO format(' The# of Radiation B.C.=' ,i6,' > Lq=' ,i6/'
* ' Change the statement to Lq=' ,i6,' and re-run')
stop
end if
c
c other minor modification
c
id(mp,np)='e'
id(1,1)='e'
id(1,np)='e'
id(mp,1)='e'
c
header=' I.D. code for each grid point'
write(8,' (a30) ') header
write(8,60) np
60 format(' j i = 1, ... ,',iS)
do j=1,np
jj=np-j+1
write ( 8 , 7 0 ) J J , ( i d ( i , j j ) , i =1, mp)
70 format(i4,1x,130a1)
end do
c
header='Given B.C. at following points'
write(8,' (iS,2x,a30) ') nobc, header
write ( 8, 8 0)
80 format(' n i j B.V.')
do i=1,nobc
write(8,90) i,ibc(i) ,jbc(i) ,zp(i)
end do
90 format(3iS,2f12.6)
c
close(8)
stop
end

III-3
Appendix IV. Source List of MATLAB program, P_PLOT.M
% Program phi_plt.M
% MathLab M file for plotting the output P matrix generated
% It uses MathLab ver. 4.2C.1 for WINDOWS

~Program developed by Jerome P.-Y. Maa at


% Virginia Institute of Marine Science, Feb. 1996
% clear the command window first.

clc;
disp('Available files for reading are : ');
disp(' 1. p bs.out; 2. p_bm.out 3. p_bL.out');
disp(' 4 p=ts.out; 5. p_tm.out 6. p_tL.out');
disp(' 0. exit ');
optlon=input('Select a number : ');

% select an input file. For other users, please change the


% fopen commend to have a proper path specification.

while option
if option == 1
[fid, message]=fopen('d:\break\p_bs.out');
end
if option == 2
[fid, message]=fopen('d:\break\p_bm.out');
end
if option == 3
[fid, message]=fopen('d:\break\p_bL.out');
end
if option == 4
(fid, message]=fopen('d:\break\p_ts.out');
end
if option == 5
(fid, message]=fopen('d:\break\p_tm.out');
end
if option == 6
(fid, message]=fopen('d:\break\p_tL.out');
end
%
% contour values for water depth

v_dep=(O 0.1 0.2 0.4 0.8 1.0 1.2 1.4 1.6 1.8 2];

% read data
ftitle=fgetl(fid);
disp (I I) j
disp(ftitle);
mp=fscanf(fid, '%d', 1);
np=fscanf(fid, '%d', 1);

IV-1
disp(['m= ',int2str(mp) ,' n= int2str(np)] ) ;
dx=fscanf(fid, '%f', 1);
dy=fscanf(fid, '%f', 1);
disp( ['dx= ',num2str(dx),' dy= ' num2str(dy)] ) ;

% read solution

p= [] ;
for ii=1:np
nk=fscanf(fid, '%d', 1);
if nk -= ii
disp(['Seq. error, i=' ,int2str(nk), '<> ii' ,int2str(ii)]);
disp ('paused') ;
pause;
end
a=fscanf(fid,'%f', mp);
p ( i i , 1 : mp) =a' ;
end
clear a;
fclose(fid);

xmin=O;
xmax=3.1415926;
ymin=O;
ymax=xmax;

% Plot the contours

elf;
axes('position', [0.1 0.1 0.88 0.88]);
[mmm nnn]=size(p);
dxx=(xmax-xmin)/(nnn-1);
dyy=(ymax-ymin)/(mmm-1);
x=xmin:dxx:xmax;
y=ymin:dyy:ymax;
axis ratio=(xmax-xmin)/(ymax-ymin);
data-ratio=1;
cl=contour(x, y, p, v_dep);
set(gca,'aspectratio', [axis_ratio, data ratio]);
set(gca,'xlim', [xmin,xmax], 'ylim', [ymin~ymax]);
hold on;
clabel(cl, 'manual');
xlabel('X '); ylabel('Y ');
title('Band Matrix Solution');

disp('Available files for reading are : ');


disp(' 1. p bs.out; 2. p_bm.out; 3. p bL.out;');
disp(' 4. p-ts.out; 5. p_tm.out; 6. p=tL.out;');
disp(' 0. exit ');
option=input('Select a number : ' ) ;
end

IV-2
Appendix V. Source listing of P_EXACT.M
,9.,
0

% This program calculate the exact solution of a Laplace equation


% with the following boundary conditions:
% at x=O, p = 1; at x=pi, p = o
% at y=O, p = 2; at y=pi, p = 0
%
% The analytical solution is provided by Jerome P.-Y. Maa.
%
% the results are stored in a ASCII file for later plot with the
% solution using a banded matrix solver, and a Thrifty banded
% matrix solver. It also plots the results at end.
%
%Because the large value of cosh(x), tanh(x), etc., the solution
% given in the text was manipulated to the form given in this
% program.
%
% one may change the statement of no=xxxx to another number to
% see the difference.
%
clear;

no=4000;

dx=0.031415926;
dy=dx;
m=(pi/dx)+l;
y=O:dy:pi; x=y; y=y';

z=zeros(m,m);

for n=l:no
ff=l-2*(-l)"'n;
aa=exp(-2*n*pi);
bb=exp(-n*x);
cc=exp(n*(x-2*pi));
dd=exp(n*(x-pi));
ee=exp(-n*(x+pi));
gg=sin(n*y);
z=z+2/(n*pi)*gg*((ff*(bb-cc)-dd+ee)/(1-aa)+l);
end
fid=fopen('c:\break\P_exact.dat', 'w');
format short;
fprintf(fid,'~alytical solution of a Laplace Equation\n');
fprintf(fid,'w~th b.C.: p=l@ x=O; p=2@ y=O; p=O at x=y=pi\n');
for i=l:m
fprintf(fid, '%d\n', i);
a=z(i,l:m);

V-1
l l lil 1001010362
l~lil fl ili ]Ji il ~ l l i~l l]l i~li j~il
fprintf(fid, '%7.4f %7.4f %7.4f %7.4f %7.4f %7.4f %7.4f
%7.4f %7.4f %7.4f\n', a);
fprintf(fid, '\n');
end
fclose(fid);

%mesh(z);
%pause;
v=[0.01 0.1 0.2 0.3 0.4 0.5 0.6 0.8 0.9 1 1.2 1.4 1.6 1.8 1.9 2];
c=contour(x,y,z,v);
axis('square');
clabel(c,'manual');

xlabel ('X');
ylabel('Y');

text(2.6, 2.9, [ 1 n=' num2str(no)], 'sc'),


end

VIMS SH 1 V48 no.135

, Maa, • - Y • Gauss~an
UsingP the .
, elimination roetho~ for
J'·
large banded matr~x

VIMS SH 1 V48 no.135


Maa, P. -Y.
Using the Gaussian
elimination method for
large banded matrix

OEMCO

V-2

You might also like