1D-FEM in Fortran: Steady State

Heat Conduction

Kengo Nakajima
Information Technology Center

Programming for Parallel Computing (616-2057)

Seminar on Advanced Computing (616-4009)
FEM1D 2

• 1D-code for Static Linear-Elastic Problems by

Galerkin FEM
• Sparse Linear Solver
– Conjugate Gradient Method
– Preconditioning
• Storage of Sparse Matrices

• Program
FEM1D 3

Keywords
• 1D Steady State Heat Conduction Problems
• Galerkin Method
• Linear Element
• Preconditioned Conjugate Gradient Method
FEM1D 4

1D Steady State Heat Conduction

heat generation/volume Q
  T  
 Q  0
x  x 

x=0 (xmin) x= xmax

• Uniform: Sectional Area: A, Thermal Conductivity: 

• Heat Generation Rate/Volume/Time [QL-3T-1] Q
• Boundary Conditions
– x=0 ： T= 0 (Fixed Temperature)
T
– x=xmax：  0 (Insulated)
x
FEM1D 5

1D Steady State Heat Conduction

heat generation/volume Q
  T  
 Q  0
x  x 

x=0 (xmin) x= xmax

• Uniform: Sectional Area: A, Thermal Conductivity: 

• Heat Generation Rate/Volume/Time [QL-3T-1] Q
• Boundary Conditions
– x=0 ： T= 0 (Fixed Temperature)
T
– x=xmax：  0 (Insulated)
x
FEM1D 6

Analytical Solution
heat generation/volume Q   T  
 Q  0
x  x 

x=0 (xmin) x= xmax

T  0@ x  0 T
 0 @ x  xmax
x
T   Q
T   Q x  C1  C1  Q xmax , T   0 @ x  xmax
1
T   Q x 2  C1 x  C2  C2  0, T  0 @ x  0
2
1  2 Q xmax
 T  Qx  x
2 
FEM1D 7

1D Linear Element (1/4)

一次元線形要素
• 1D Linear Element
– Length= L
• Node (Vertex)
• Element

– Ti Temperature at i
T Ti
– Tj Temperature at j
– Temperature T on each
Tj
element is linear
function of x (Piecewise
Linear): L

x
T  1   2 x Xi Xj
FEM1D 8

1D Linear Element (1/4)

一次元線形要素
• 1D Linear Element
– Length= L
• Node (Vertex)
• Element

– Ti Temperature at i
T Ti
– Tj Temperature at j
– Temperature T on each
Tj
element is linear
function of x (Piecewise
Linear): L

x
T  1   2 x Xi Xj
FEM1D 9

Piecewise Linear

1 2 3 4 5 Global Node ID
1 2 3 4 Element ID

1 2 1 2 1 2 1 2
1 2 3 4

Local Node ID
for each elem.
Gradient of temperature is constant in each
element (might be discontinuous at each “node”)
FEM1D 10

1D Linear Elem.: Shape Function (2/4)

• Coef’s are calculated basedT Ti
on info. at each node
T  Ti @ x  X i , T  T j @ x  X j Tj
Ti  1   2 X i , T j  1   2 X j
L
• Coefficients:
Ti X j  T j X i T j  Ti x
Xi Xj
1  , 2 
L L
• T can be written as follows, according to Ti and Tj :
Ni, Nj
 Xj x  x  Xi 
T    Ti    Tj Shape Function or
 L   L 
Interpolation Function
Ni Nｊ function of x (only)
FEM1D 11

1D Linear Elem.: Shape Function (3/4)

• Number of Shape Functions T
T i
= Number of Vertices of
Each Element Tj
– Ni: Function of Position
– A kind of Test/Trial Functions L
 Xj x  x  Xi 
N i   , N j    Xi Xj
x
 L   L 
• Linear combination of shape functions provides
displacement “in” each element
– Coef’s (unknows): Temperature at each node
i Trial/Test Function (known
M function of position, defined in
T  N iTi  N jT j TuMM   ai i domain and at boundary. “Basis”
i 1 in linear algebra.
ai Coefficients (unknown)
FEM1D 12

1D Linear Elem.: Shape Function (4/4)

T Ti
• Value of Ni
– =1 at one of the nodes in
Tj
element
– =0 on other nodes L
 Xj x  x  Xi 
N i   , N j    x
 L   L  Xi Xj
N

1.0 Ni Nj

x
Xi Xj
FEM1D 13

Galerkin Method (1/4)

• Governing Equation for 1D
Steady State Heat
Conduction Problems
(Uniform ):
 d 2T  
  2   Q  0
 dx 
T  N {} Distribution of temperature in each element
(matrix form), : Temperature at each node
• Following integral equation is obtained at each
element by Galerkin method, where [N]’s are also
weighting functions:
  d 2T   
V N    dx 2   QdV  0
T

 
FEM1D 14

Galerkin Method (2/4)

体積当たり一様発熱 Q
• Green’s Theorem (1D)
 d 2B  dB  dA dB 
x=0 (xmin) x= xmax

V A dx 2  dV  S A dx dS V  dx dx  dV

• Apply this to the 1st part of eqn with 2nd-order diff.:
 d 2T   d N T dT 
   T dT
T
V  N  
 dx 2 
 
dV    
V 

 dx dx 


dV  
S
 N
dx
dS

• Consider the following terms: V

dT d N  S S
T  N {},  {} q   dT
dx dx dx
: Heat flux at element surface[QL-2T-1]
FEM1D 15

Galerkin Method (3/4)

• Finally, following eqn is 体積当たり一様発熱 Q

obtained by considering
heat generation term Q ： x=0 (xmin) x= xmax

 d N T d N  
    dV    V
V 
dx dx  S S

  q N  dS   QN  dV  0
T T

S V

• This is called “weak form（弱形式）”. Original PDE

consists of terms with 2nd-order diff., but this “weak
form” only includes 1st-order diff by Green’s theorem.
– Requirements for shape functions are “weaker” in “weak
form”. Linear functions can describe effects of 2nd-order
differentiation.
FEM1D 16

Galerkin Method (4/4)

体積当たり一様発熱 Q
 d N T d N  
    dV   
V 
dx dx  x=0 (xmin) x= xmax

 q N  dS  Q N  dV  0
 
T T
V
S V
S S

• These terms coincide at element boundaries and

disappear. Finally, only terms on the domain
boundaries remain.
FEM1D 17

Weak Form and Boundary Conditions

体積当たり一様発熱 Q
• Value of dependent variable is
defined (Dirichlet)
x=0 (xmin) x= xmax
– Weighting Function = 0
– Principal B.C. (Boundary V
Condition)（第一種境界条件） S S
– Essential B.C.（基本境界条件）

• Derivatives of Unknowns  d N T d N  
    dV   
(Neumann) 
V  
dx dx
– Naturally satisfied in weak form  q N  dS  Q N  dV  0
 
T T

– Secondary B.C.（第二種境界 S V

条件） dT
where q  
– Natural B.C（自然境界条件） dx
FEM1D 18

Weak Form with B.C.: on each elem.

k   
(e) (e)
 f 
(e)

 d N  d N  
T
k  (e)
    dV

V  
dx dx

f  (e)
  QN  dV   q N  dS
 T T

V S
FEM1D 19

Integration over Each Element: [k]

 Xj x  x  Xi  dN i   1  1
dN j
N i   , N j      ,  
 L   L  dx  L  dx  L 
 d N T d N   ui
V   dx dx dV u

L
 1 / L  uj
     1 / L, 1 / L A dx
0 
1/ L 
L
2x1 matrix 1x2 matrix
x
A L  1  1 A  1  1 Xi  0 Xj L

L2   1 
 1
dx 
L  1  1
0
 x x
N i  1  , N j   
A: Sectional Area  L L
L: Length
FEM1D 20

Integration over Each Element: {f} (1/2)

 Xj x  x  Xi  dN i   1  1
dN j

Ni   
, Nj     ,  
 L   L  dx  L  dx  L 

 x x
N i  1  , N j   
 L L

 N  dV  Q A
L
1  x / L  Q AL 1 Heat Generation
V 0  x / L  dx  2 1
T
Q (Volume)

1 : 1
A: Sectional Area
L: Length
FEM1D 21

Integration over Each Element: {f} (2/2)

 Xj x  x  Xi  dN i   1  dN j1

Ni   
, Nj     ,  
 L   L  dx  L  dx  L 

L
1  x / L   AL 1 Heat Generation
V Q N  dV  Q A0  x / L  dx  2 1
T Q
(Volume)

0
 q N  dS  q A x L
dT Surface Heat Flux
 q A , q  
T

S 1 dx

when surface heat flux acts on

only this surface.
FEM1D 22

Global Equations
• Accumulate Element Equations:

k (e)  (e)   f (e) Element Matrix, Element Equations

K    F  Global Matrix, Global Equations

K    k , F     f 
: global vector of  
This is the final linear equations
(global equations) to be solved.
FEM1D 23

ECCS2012 System

Creating Directory
>$ cd Documents
>$ mkdir 2013summer your favorite name
>$ cd 2013summer

This is your “top” directory, and is called <$E-TOP> in this class.

1D Code for Steady-State Heat Conduction Problems

>$ cd <$E-TOP>
>$ cp /home03/skengon/Documents/class_eps/F/1d.tar .
>$ cp /home03/skengon/Documents/class_eps/C/1d.tar .
>$ tar xvf 1d.tar
>$ cd 1d
FEM1D 24

Compile & GO !
>$ cd <$E-TOP>/1d
>$ cc –O 1d.c (or g95 –O 1d.f)
>$ ./a.out

Control Data input.dat

4 NE (Number of Elements)
1.0 1.0 1.0 1.0 x (Length of Each Elem.: L) ，Q, A, 
100 Number of MAX. Iterations for CG Solver
1.e-8 Convergence Criteria for CG Solver

x=1
Element ID
1 2 3 4 5
Node ID (Global)
1 2 3 4
x=0 x=1 x=2 x=3 x=4
FEM1D 25

Results
>$ ./a.out

4 iters, RESID= 4.154074e-17

### TEMPERATURE
1 0.000000E+00 0.000000E+00
2 3.500000E+00 3.500000E+00
3 6.000000E+00 6.000000E+00
4 7.500000E+00 7.500000E+00
5 8.000000E+00 8.000000E+00

Computational Analytical

x=1
Element ID
1 2 3 4 5
Node ID (Global)
1 2 3 4
x=0 x=1 x=2 x=3 x=4
FEM1D 26

Element Eqn’s/Accumulation (1/3)

• 4 elements, 5 nodes
Element ID
1 2 3 4 5
Node ID (Global)
1 2 3 4
x=0 x=1 x=2 x=3 x=4

• [k] and {f} of Element-1:

A  1  1 Q AL 1
k  (1)
 f  (1)
 
L   1  1 2 1

• As for Element-4:
A  1  1 Q AL 1
k 
( 4)
 f  ( 4)
 
L   1  1 2 1
FEM1D 27

Element Eqn’s/Accumulation (2/3)

• Element-by-Element Accumulation:

4
K    k (e)  + + +
e 1

4
F     f (e)  + + +
e 1
FEM1D 28

Element Eqn’s/Accumulation (3/3)

• Relations to FDM
A  1  1
k  (e)

L  1  1

4
+1 -1
A
K    k  [ ] L
(e) -1 +1 +1 -1
 + -1 +1 + +1 -1 +
e 1 -1 +1 +1 -1
-1 +1

+1 -1  d 2T   T  2Ti  Ti 1 
-1 +2 -1 A    2 dV     i 1 dV

2
 
-1 +2 -1  V
dx  V
L
L  T  2Ti  Ti 1 
  AL  Ti 1  2Ti  Ti 1  
-1 +2 -1 A
-1 +1
  i 1 2
 L  L
Something familiar …
FEM: Coefficient Matrices are generally sparse
(many ZERO’s)
FEM1D 29

2nd –Order Differentiation in FDM

• Approximate Derivative at×(center of i and i+1)

i-1 i i+1
 d  i1  i
×   
x x  dx i 1/ 2 x

x→0: Real Derivative

• 2nd-Order Diff. at i

 d   d  i1  i i  i1
    
 d 2 
 2  
 dx i 1/ 2  dx i1/ 2
 x x  i 1  2i  i 1
 dx i x x x2
FEM1D 30

Element-by-Element Operation
very flexible if each element has different material
property, size, etc.

( e ) A( e )  1  1
k  (e)
   1  1
L( e )  

+1 -1

K    k  (1) A(1) ( 2) A( 2)
4
(e) -1 +1 +1 -1
 (1)
+ -1 +1 
e 1 L L( 2 )

(3) A(3) ( 4) A( 4)
+1 -1  ( 3)
+ 
-1 +1 L +1 -1 L( 4 )
-1 +1
FEM1D 31

Element/Global Operations
1 2 3 4
1 2 3 4 5 Global Node ID

Local Node ID
e e e e
1 2 1 2 1 2 1 2
 d N  d N  
T
 d N  d N  
T
 d N  d N  
T
 d N  d N  
T
  E  dV      E  dV      E  dV      E  dV   
V  dx dx  V  dx dx  V  dx dx  V  dx dx 
   N  dS   X N  dV  0    N  dS   X N  dV  0    N  dS   X N  dV  0    N  dS   X N  dV  0
T T T T T T T T

S V S V S V S V

[k ]( 3) {}( 3)  { f }(3) [k ]( 4 ) {}( 4)  { f }( 4 )

[k ](1) {}(1)  { f }(1) [k ]( 2 ) {}( 2 )  { f }( 2 )
k11( 3) k12( 3)  1( 3)   f1( 3)  k11( 4 ) k12( 4 )  1( 4 )   f1( 4) 
k11(1) k12(1)  1(1)   f1(1)  k11( 2 ) k12( 2 )  1( 2)   f1( 2)    (4) 
  ( 2)   ( 3) ( 3)   ( 3) 
  ( 3)   ( 4) ( 4)   ( 4) 
 (1) (1)   (1) 
  (1)   (2) ( 2)   (2)   2   f 2 
 2   f 2  k 21 k 22
 2   f 2  k 21 k 22
 2   f 2  k 21 k 22
k 21 k 22

1 2 3 4
1 2 3 4 5
[ K ]{}  {F }
 D1 AU11   1   B1 
Mapping Information needed,
 AL D2 AU 21     B  from element-matrix
 21   2   2 
 AL31 D3 AU 31   3    B3  to global-matrix.
 
 AL41 D4 AU 41   4   B4 
   
 AL51 D5   5   B5 
FEM1D 32
Around the FEM in 3 minutes

Accumulation to
Global/overall Matrix
[K ]{}  {F}
13 14 15 16  D X X X  1   F1 
X D X X
X X     F 
7 8 9   2   2 
 X D X X X X  3   F3 
    
9 10 11 12  X D X X  4   F4 
X X D X X X  5   F5 
4 5 6     
X X X X D X X X X  6   F6 
5 6 7 8  X X X X D X X X X     F 
  7   7 
 X X X D X X  8   F8 
1 2 3  X X D X X X      F 
  9   9 
1 2 3 4  X X X X D X X X X 10  F10 
    
 X X X X D X X X X 11  F11 
 X X X D X X 12  F12 
    
 X X D X 13  F13 
 X X X X D X 14  F 
    14 
 X X X X D X 15  F15 
    
 X X X D16  F16 
FEM1D 33

• 1D-code for Static Linear-Elastic Problems by

Galerkin FEM
• Sparse Linear Solver
– Conjugate Gradient Method
– Preconditioning
• Storage of Sparse Matrices

• Program
FEM1D 34

Large-Scale Linear Equations in

Scientific Applications
• Solving large-scale linear equations Ax=b is the most
important and expensive part of various types of
scientific computing.
– for both linear and nonlinear applications
• Various types of methods proposed & developed.
– for dense and sparse matrices
– classified into direct and iterative methods
• Dense Matrices:密行列: Globally Coupled Problems
– BEM, Spectral Methods, MO/MD (gas, liquid)
• Sparse Matrices:疎行列: Locally Defined Problems
– FEM, FDM, DEM, MD (solid), BEM w/FMM
FEM1D

Direct Method
直接法
 Gaussian Elimination/LU Factorization.
 compute A-1 directly.
Good
 Robust for wide range of applications.
 Good for both dense and sparse matrices
Bad
 More expensive than iterative methods (memory, CPU)
 not scalable

35
FEM1D

Iterative Method
反復法
• Stationary Method
– SOR, Gauss-Seidel, Jacobi
– Generally slow, impractical

• Non-Stationary Method
– With restriction/optimization conditions
– Krylov-Subspace
– CG: Conjugate Gradient
– BiCGSTAB: Bi-Conjugate Gradient Stabilized
– GMRES: Generalized Minimal Residual
36
FEM1D

Iterative Method (cont.)

Good
 Less expensive than direct methods, especially in memory.
 Suitable for parallel and vector computing.

Bad
 Convergence strongly depends on problems, boundary
conditions (condition number etc.)
 Preconditioning is required : Key Technology for
Parallel FEM

37
FEM1D 38

Conjugate Gradient Method

共役勾配法
• Conjugate Gradient: CG
– Most popular “non-stationary” iterative method
• for Symmetric Positive Definite (SPD) Matrices
– 対称正定
– {x}T[A]{x}>0 for arbitrary {x}
– All of diagonal components, eigenvaules and leading
principal minors > 0 （主小行列式・首座行列式）
– Matrices of Galerkin-based FEM: heat conduction, Poisson,
static linear elastic problems a a a a  a 
11 12 13 14 1n
a a24  a2 n 
• Algorithm  21
a
a22
a32
a23
a33 a34  a3n 
det  31 
– “Steepest Descent Method” a41 a42 a43 a44  a4 n 
   
– x(i)= x(i-1) + i p(i) 
an1

an 2

an 3


an 4  ann 
• x(i)：solution，p(i)：search direction，i: coefficient
– Solution {x} minimizes {x-y}T[A]{x-y}, where {y} is exact
solution.
FEM1D 39

Procedures of Conjugate Gradient

Compute r(0)= b-[A]x(0)
for i= 1, 2, … • Mat-Vec. Multiplication
z(i-1)= r(i-1) • Dot Products
i-1= r(i-1) z(i-1)
if i=1 • DAXPY (Double
p(1)= z(0)
else
Precision: a{X} + {Y})
i-1= i-1/i-2
p(i)= z(i-1) + i-1 p(i-1)
endif
q(i)= [A]p(i) x(i) : Vector
i = i-1/p(i)q(i)
x(i)= x(i-1) + ip(i) i : Scalar
r(i)= r(i-1) - iq(i)
check convergence |r|
end
FEM1D 40

Procedures of Conjugate Gradient

Compute r(0)= b-[A]x(0)
for i= 1, 2, … • Mat-Vec. Multiplication
z(i-1)= r(i-1) • Dot Products
i-1= r(i-1) z(i-1)
if i=1 • DAXPY
p(1)= z(0)
else
i-1= i-1/i-2
p(i)= z(i-1) + i-1 p(i-1)
endif
q(i)= [A]p(i)
i = i-1/p(i)q(i)
x(i) : Vector
x(i)= x(i-1) + ip(i) i : Scalar
r(i)= r(i-1) - iq(i)
check convergence |r|
end
FEM1D 41

Procedures of Conjugate Gradient

Compute r(0)= b-[A]x(0)
for i= 1, 2, … • Mat-Vec. Multiplication
z(i-1)= r(i-1) • Dot Products
i-1= r(i-1) z(i-1)
if i=1 • DAXPY
p(1)= z(0)
else
i-1= i-1/i-2
p(i)= z(i-1) + i-1 p(i-1)
endif
q(i)= [A]p(i) x(i) : Vector
i = i-1/p(i)q(i)
x(i)= x(i-1) + ip(i) i : Scalar
r(i)= r(i-1) - iq(i)
check convergence |r|
end
FEM1D 42

Procedures of Conjugate Gradient

Compute r(0)= b-[A]x(0)
for i= 1, 2, … • Mat-Vec. Multiplication
z(i-1)= r(i-1) • Dot Products
i-1= r(i-1) z(i-1)
if i=1 • DAXPY
p(1)= z(0)
else
• Double
i-1= i-1/i-2 • {y}= a{x} + {y}
p(i)= z(i-1) + i-1 p(i-1)
endif
q(i)= [A]p(i) x(i) : Vector
i = i-1/p(i)q(i)
x(i)= x(i-1) + ip(i) i : Scalar
r(i)= r(i-1) - iq(i)
check convergence |r|
end
FEM1D 43

Procedures of Conjugate Gradient

Compute r(0)= b-[A]x(0)
for i= 1, 2, …
x(i) : Vector
z(i-1)= r(i-1) i : Scalar
i-1= r(i-1) z(i-1)
if i=1
p(1)= z(0)
else
i-1= i-1/i-2
p(i)= z(i-1) + i-1 p(i-1)
endif
q(i)= [A]p(i)
i = i-1/p(i)q(i)
x(i)= x(i-1) + ip(i)
r(i)= r(i-1) - iq(i)
check convergence |r|
end
FEM1D 44

Algorithm of CG Method (1/5)

Solution x minimizes the following equation if y is the exact
solution (Ay=b)
x  y T Ax  y 
x  y T Ax  y   x, Ax    y, Ax   x, Ay    y, Ay 
 x, Ax   2 x, Ay    y, Ay    x, Ax   2x, b    y, b  Const.

Therefore, the solution x minimizes the following f(x):

1
f x    x, Ax    x, b 
2
1
f x  h   f  x   h, Ax  b   h, Ah  Arbitrary vector h
2
FEM1D 45

1
f x    x, Ax   x, b 
2
1
f x  h   f  x   h, Ax  b   h, Ah  Arbitrary vector h
2

1
f  x  h    x  h, A( x  h)    x  h, b 
2
1 1
  x  h, Ax   x  h, Ah    x, b   h, b 
2 2
1 1 1 1
  x, Ax   h, Ax    x, Ah   h, Ah    x, b   h, b 
2 2 2 2
1 1
  x, Ax    x, b   h, Ax   h, b   h, Ah 
2 2
1
 f  x   h, Ax  b   h, Ah 
2
FEM1D 46

Algorithm of CG Method (2/5)

CG method minimizes f(x) at each iteration.
Assume that approximate solution: x(0), and
search direction vector p(k) is defined at k-th iteration.
x ( k 1)  x ( k )   k p ( k )

Minimization of f(x(k+1)) is done as follows:

 (k ) (k )
1 2 (k )
2
     
f x   k p   k p , Ap ( k )   k p ( k ) , b  Ax ( k )  f x ( k )


f x ( k )   k p ( k ) 
 0  k 
  
p ( k ) , b  Ax ( k ) 
p(k ) , r (k )
 (k )
 k 
(k )
p , Ap  
(k )

p , Ap ( k )

r ( k )  b  Ax ( k ) residual vector
FEM1D 47

Algorithm of CG Method (3/5)

Residual vector at (k+1)-th iteration: r ( k 1)  b  Ax ( k 1) , r ( k )  b  Ax ( k )

r ( k 1)
r (k )
  k Ap (k ) r ( k 1)  r ( k )  Ax ( k 1)  Ax ( k )   k Ap ( k )

Search direction vector p is defined by the following

recurrence formula:
p ( k 1)  r ( k 1)   k p ( k ) , r ( 0 )  p ( 0 )

It’s lucky if we can get exact solution y at (k+1)-th iteration:

y  x ( k 1)   k 1 p ( k 1)
FEM1D 48

Algorithm of CG Method (4/5)

BTW, we have the following (convenient) orthogonality relation:
Ap (k )
, y  x ( k 1)   0

Ap (k )
   
, y  x ( k 1)  p ( k ) , Ay  Ax ( k 1)  p ( k ) , b  Ax ( k 1) 
 p (k )
  
, b  A x ( k )   k p ( k )  p ( k ) , b  Ax ( k )   k Ap ( k ) 
 p (k )
    
, r ( k )   k Ap ( k )  p ( k ) , r ( k )   k p ( k ) , Ap ( k )  0

 k 
 p (k )
, r (k )
p (k )
, Ap ( k ) 
Thus, following relation is obtained:
Ap (k )
   
, y  x ( k 1)  Ap ( k ) ,  k 1 p ( k 1)  0  p ( k 1) , Ap ( k )  0 
FEM1D 49

Algorithm of CG Method (5/5)

p ( k 1)
, Ap   r
(k )
  p , Ap   r
( k 1)
k , Ap     p , Ap   0
(k ) (k ) ( k 1) (k )
k
(k ) (k )

 r , Ap 
( k 1) (k )
 k 
 p , Ap 
(k ) (k )

p ( k 1)
, Ap   0 p(k) is “conjugate” for matrix A
(k )

Following “conjugate” relationship is obtained for arbitrary (i,j):

p (i )
, Ap ( j )   0 i  j 
Following relationships are also obtained for p(k) and r(k):
r (i )
, r ( j )   0 i  j , p (k )
, r ( k )   r ( k ) , r ( k ) 
In N-dimensional space, only N sets of orthogonal and linearly
independent residual vector r(k). This means CG method
converges after N iterations if number of unknowns is N.
Actually, round-off error sometimes affects convergence.
FEM1D 50

Proof (1/2)
p (i )
, r ( k 1)   0, i  0,1, , k
k
x ( k 1)  x ( i 1)   j
 p
j i 1
( j)

 (i 1) k 
r ( k 1)
 b  Ax ( k 1)
 b  A x   j p ( j)

 j i 1 

  
k k
 b  Ax (i 1) 
j i 1
j Ap ( j)
 r ( i 1)
  j
 Ap
j i 1
( j)

 ( i ) ( i 1) 
p    p , r    j Ap 
k
(i ) ( k 1) ( j)
,r
 j i 1 
Ap , y  x   0
(k ) ( k 1)

 (i ) k 
Ap , y  x 
(k ) ( k 1)

  p , r    p ,   j Ap   0
(i ) ( i 1) ( j)
  p , Ay  Ax
(k )
 ( k 1)

 j i 1    p , b  Ax
(k )
 ( k 1)

 p , r (k )
 0
( k 1)

=0 =0
FEM1D 51

Proof (2/2)
r (i )
, r ( j )   0 i  j 

  
0  p ( i ) , r ( k 1)  r ( i )   i 1 p (i 1) , r ( k 1) 
  i 1 p ( i 1)
, r ( k 1)
 
 r (i )
, r ( k 1)
 
 r (i )
, r ( k 1)

p (k )
, r ( k )   r ( k ) , r ( k ) 

p (k )
, r ( k )   r ( k )   k 1 p ( k 1) , r ( k ) 
  k 1 p ( k 1) , r ( k )   r ( k ) , r ( k )   r ( k ) , r ( k ) 
FEM1D 52

k，k
Usually, we use simpler definitions of k，k as follows:

 
 p , b  Ax   p
(k )

(k ) (k )
, r (k )  r (k ) , r (k ) 
 (k )
k
 p , Ap   p
(k ) (k ) (k )
, Ap   p , Ap ( k ) 
(k )

  p , r   r , r 
(k ) (k ) (k ) (k )

 r ( k 1) , Ap ( k )  r ( k 1) , r ( k 1) 

k  
 p , Ap  r ( k ) , r ( k ) 
(k ) (k )

r  ( k 1)
, Ap (k )
 r ( k 1)
, r ( k )  r ( k 1) 

r ( k 1)
, r ( k 1) 
k k
FEM1D 53

Procedures of Conjugate Gradient

Compute r(0)= b-[A]x(0)
for i= 1, 2, …
x(i) : Vector
z(i-1)= r(i-1) i : Scalar
i-1= r(i-1) z(i-1)
if i=1
p(1)= z(0)
else
i-1= i-1/i-2
p(i)= z(i-1) + i-1 p(i-1)
endif
q(i)= [A]p(i)
i = i-1/p(i)q(i)
x(i)= x(i-1) + ip(i)
r(i)= r(i-1) - iq(i)
check convergence |r|
end
FEM1D 54

Preconditioning for Iterative Solvers

 Convergence rate of iterative solvers strongly depends
on the spectral properties (eigenvalue distribution) of
the coefficient matrix A.
 Eigenvalue distribution is small, eigenvalues are close to 1
 In "ill-conditioned" problems, "condition number" (ratio of
max/min eigenvalue if A is symmetric) is large.
 A preconditioner M (whose properties are similar to
those of A)transforms the linear system into one with
more favorable spectral properties
 In "ill-conditioned" problems, "condition number" (ratio of
max/min eigenvalue if A is symmetric) is large.
 M transforms original equation Ax=b into A'x=b' where
A'=M-1A, b'=M-1b
 If M~A, M-1A is close to identity matrix.
 If M-1=A-1, this is the best preconditioner (a.k.a. Gaussian
Elimination)
FEM1D 55

Preconditioned CG Solver
Compute r(0)= b-[A]x(0)
for i= 1, 2, …
solve [M]z(i-1)= r(i-1)
i-1= r(i-1) z(i-1)
if i=1
p(1)= z(0)
else x

i-1= i-1/i-2
p(i)= z(i-1) + i-1 p(i-1)
endif
q(i)= [A]p(i)
i = i-1/p(i)q(i)
x(i)= x(i-1) + ip(i)
r(i)= r(i-1) - iq(i)
check convergence |r|
end
FEM1D 56

ILU(0), IC(0)
• Widely used Preconditioners for Sparse Matrices
– Incomplete LU Factorization
– Incomplete Cholesky Factorization (for Symmetric
Matrices)

• Incomplete Direct Method

– Even if original matrix is sparse, inverse matrix is not
necessarily sparse.
– fill-in
– ILU(0)/IC(0) without fill-in have same non-zero pattern
with the original (sparse) matrices
FEM1D 57

Diagonal Scaling, Point-Jacobi

D1 0 ... 0 0
0 D 0 0 
 2 
M    ... ... ... 
 
0 0 DN 1 0 
 0 0 ... 0 DN 

• solve [M]z(i-1)= r(i-1) is very easy.

• Provides fast convergence for simple problems.
• 1d.f, 1d.c
FEM1D 58

• More detailed discussions on preconditioning will be

provided in “Multicore Programming”.
FEM1D 59

• 1D-code for Static Linear-Elastic Problems by

Galerkin FEM
• Sparse Linear Solver
– Conjugate Gradient Method
– Preconditioning
• Storage of Sparse Matrices

• Program
FEM1D 60

Coef. Matrix derived from FEM

D X X X  1   F1 
• Sparse Matrix X


D X
X D X
X X X
X X X
    F 
 2   2 
 3   F3 
    
– Many “0”’s  X D X X  4   F4 
X X D X X X  5   F5 
    
X X X X D X X X X  6   F6 
• Storing all components 

X X X X D X X X X     F 
 7   7 
 X X X D X X  8   F8 
(e.g. A(i,j)) is not efficient 


X X
X X X
D X
X D X
X X
X X X
     F 
 9   9 
10  F10 
    
for sparse matrices 

X X X
X X
X D X
X D
X X X 11  F11 
X X 12  F12 
    
– A(i,j) is suitable for dense 

X X
X X X
D X
X D X
13  F13 
14  F14 
    
X D X 15  F15 
matrices 


X X X
X X
   
X D16  F16 

• Number of non-zero off-diagonal components is

O(100) in FEM
– If number of unknowns is 108 :
• A(i,j): O(1016) words
• Actual Non-zero Components：O(1010) words
• Only (really) non-zero off-diag. components should
be stored on memory
FEM1D 61

Variables/Arrays in 1d.f, 1d.c

related to coefficient matrix
name type size description
N I - # Unknowns

NPLU I - # Non-Zero Off-Diagonal Components

Diag(:) R N Diagonal Components
U(:) R N Unknown Vector

Rhs(:) R N RHS Vector

0:N Off-Diagonal Components (Number of Non-Zero Off-
Index(:) I
N+1 Diagonals at Each ROW)
Item(:) I NPLU Off-Diagonal Components (Corresponding Column ID)

AMat(:) R NPLU Off-Diagonal Components (Value)

Only non-zero components are stored according

to “Compressed Row Storage”.
FEM1D 62

Mat-Vec. Multiplication for Sparse Matrix

Compressed Row Storage (CRS)
Diag (i) Diagonal Components (REAL, i=1~N)
Index(i) Number of Non-Zero Off-Diagonals at Each ROW (INT, i=0~N)
Item(k) Off-Diagonal Components (Corresponding Column ID)
（INT, k=1, index(N)）
AMat(k) Off-Diagonal Components (Value)
（ REAL, k=1, index(N) ）

{Y}= [A]{X} D
X
X X X  1   F1 
    F 
 D X X X X  2   2 
 X D X X X X  3   F3 
    
 4   F4 
do i= 1, N  X D X X
X X D X X X  5   F5 
    
X X X X D X X X X  6   F6 
Y(i)= Diag(i)*X(i) 


X X X
X X
X D X
X D
X X X
X X
    F 
 7   7 
 8   F8 

do k= Index(i-1)+1, Index(i) 


X X
X X X
D X
X D X
X X
X X X
     F 
 9   9 
10  F10 
   
Y(i)= Y(i) + Amat(k)*X(Item(k))

 X X X X D X X X X 11  F11 
 X X X D X X 12  F12 
    

enddo 


X X
X X X
D X
X D X
13  F13 
  F 
 14   14 

enddo  X X X X D X 15  F15 

    
 X X X D16  F16 
FEM1D 63

CRS or CSR ?
for Compressed Row Storage
• In Japan and USA, “CRS” is very general for
abbreviation of “Compressed Row Storage”, but they
usually use “CSR” in Europe (especially in France).

• “CRS” in France
– Compagnie Républicaine de Sécurité
• Republic Security Company of France
• French scientists may feel
uncomfortable when we use “CRS” in
technical papers and/or presentations.
FEM1D 64

Mat-Vec. Multiplication for Sparse Matrix

Compressed Row Storage (CRS)

{Q}=[A]{P}

for(i=0;i<N;i++){
W[Q][i] = Diag[i] * W[P][i];
for(k=Index[i];k<Index[i+1];k++){
W[Q][i] += AMat[k]*W[P][Item[k]];
}
}
FEM1D 65

Mat-Vec. Multiplication for Dense Matrix

Very Easy, Straightforward
 a11 a12 ... a1, N 1 a1, N   x1   y1 
 a a a a  x   y 
 21 22 2, N 1 2, N   2   2 
 ... ... ...      
  x   y 
aN 1,1 aN 1,2 aN 1, N 1 aN 1, N 
 N 1   N 1 
 aN ,1 aN ,2 ... aN , N 1 aN , N   xN   yN 
 

{Y}= [A]{X}

do j= 1, N
Y(j)= 0.d0
do i= 1, N
Y(j)= Y(j) + A(i,j)*X(i)
enddo
enddo
FEM1D 66

Compressed Row Storage (CRS)

1 2 3 4 5 6 7 8
1 1.1 2.4 0 0 3.2 0 0 0 
2
4.3 3.6 0 2.5 0 3.7 0 9.1 

3 0 0 5.7 0 1.5 0 3.1 0 
 
4
0 4.1 0 9.8 2.5 2.7 0 0 
5 3.1 9.5 10.4 0 11.5 0 4.3 0 
 
6 0 0 6.5 0 0 12.4 9.5 0 
7 0 6.4 2.5 0 0 1.4 23.1 13.1
 
8  0 9.5 1.3 9.6 0 3.1 0 51.3
FEM1D 67

Compressed Row Storage (CRS):

Fortran
1 2 3 4 5 6 7 8
1.1 2.4 3.2 N= 8
1
① ② ⑤
4.3 3.6 2.5 3.7 9.1 対角成分
2 Diag(1)= 1.1
① ② ④ ⑥ ⑧
5.7 1.5 3.1 Diag(2)= 3.6
3 Diag(3)= 5.7
③ ⑤ ⑦
4.1 9.8 2.5 2.7 Diag(4)= 9.8
4 Diag(5)= 11.5
② ④ ⑤ ⑥
Diag(6)= 12.4
3.1 9.5 10.4 11.5 4.3
5 Diag(7)= 23.1
① ② ③ ⑤ ⑦
Diag(8)= 51.3
6.5 12.4 9.5
6
③ ⑥ ⑦
6.4 2.5 1.4 23.1 13.1
7
② ③ ⑥ ⑦ ⑧
8 9.5 1.3 9.6 3.1 51.3
② ③ ④ ⑥ ⑧
FEM1D 68

Compressed Row Storage (CRS)

1 2 3 4 5 6 7 8
1.1 2.4 3.2
1
① ② ⑤
3.6 4.3 2.5 3.7 9.1
2
② ① ④ ⑥ ⑧
5.7 1.5 3.1
3
③ ⑤ ⑦
4 9.8 4.1 2.5 2.7
④ ② ⑤ ⑥
11.5 3.1 9.5 10.4 4.3
5
⑤ ① ② ③ ⑦
12.4 6.5 9.5
6
⑥ ③ ⑦
23.1 6.4 2.5 1.4 13.1
7
⑦ ② ③ ⑥ ⑧
8 51.3 9.5 1.3 9.6 3.1
⑧ ② ③ ④ ⑥
FEM1D 69

Compressed Row Storage (CRS)

# Non-Zero
index(0)= 0
Off-Diag.
1.1 2.4 3.2
1 2 index(1)= 2
① ② ⑤
3.6 4.3 2.5 3.7 9.1 4 index(2)= 6
2
② ① ④ ⑥ ⑧
5.7 1.5 3.1
3 2 index(3)= 8
③ ⑤ ⑦
9.8 4.1 2.5 2.7
4 3 index(4)= 11
④ ② ⑤ ⑥
11.5 3.1 9.5 10.4 4.3
5 4 index(5)= 15
⑤ ① ② ③ ⑦
12.4 6.5 9.5
6 2 index(6)= 17
⑥ ③ ⑦
23.1 6.4 2.5 1.4 13.1 4 index(7)= 21
7
⑦ ② ③ ⑥ ⑧
8 51.3 9.5 1.3 9.6 3.1 4 index(8)= 25 NPLU= 25
⑧ ② ③ ④ ⑥ (=index(N))
index(i-1)+1th~index(i) th
Non-Zero Off-Diag. Components corresponding to i-th row
FEM1D 70

Compressed Row Storage (CRS)

# Non-Zero
index(0)= 0
Off-Diag.
1.1 2.4 3.2
1 2 index(1)= 2
① ②,1 ⑤,2
3.6 4.3 2.5 3.7 9.1 4 index(2)= 6
2
② ①,3 ④,4 ⑥,5 ⑧,6
5.7 1.5 3.1
3 2 index(3)= 8
③ ⑤,7 ⑦,8
4 9.8 4.1 2.5 2.7
3 index(4)= 11
④ ②,9 ⑤,10 ⑥,11
11.5 3.1 9.5 10.4 4.3
5 4 index(5)= 15
⑤ ①,12 ②,13 ③,14 ⑦,15
12.4 6.5 9.5
6 2 index(6)= 17
⑥ ③,16 ⑦,17
23.1 6.4 2.5 1.4 13.1 4 index(7)= 21
7
⑦ ②,18 ③,19 ⑥,20 ⑧,21
8 51.3 9.5 1.3 9.6 3.1 4 index(8)= 25 NPLU= 25
⑧ ②,22 ③,23 ④,24 ⑥,25 (=index(N))
index(i-1)+1th~index(i) th
Non-Zero Off-Diag. Components corresponding to i-th row
FEM1D 71

Compressed Row Storage (CRS)

1.1 2.4 3.2

1
① ②,1 ⑤,2
3.6 4.3 2.5 3.7 9.1 Example:
2 item( 7)= 5, AMAT( 7)= 1.5
② ①,3 ④,4 ⑥,5 ⑧,6
item(19)= 3, AMAT(19)= 2.5
5.7 1.5 3.1
3
③ ⑤,7 ⑦,8
4 9.8 4.1 2.5 2.7
④ ②,9 ⑤,10 ⑥,11
11.5 3.1 9.5 10.4 4.3
5
⑤ ①,12 ②,13 ③,14 ⑦,15
12.4 6.5 9.5
6
⑥ ③,16 ⑦,17
23.1 6.4 2.5 1.4 13.1
7
⑦ ②,18 ③,19 ⑥,20 ⑧,21
8 51.3 9.5 1.3 9.6 3.1
⑧ ②,22 ③,23 ④,24 ⑥,25
FEM1D 72

Compressed Row Storage (CRS)

1.1 2.4 3.2 Diag (i) Diagonal Components (REAL, i=1~N)

1
① ②,1 ⑤,2 Index(i) Number of Non-Zero Off-Diagonals at
3.6 4.3 2.5 3.7 9.1 Each ROW (INT, i=0~N)
2
② ①,3 ④,4 ⑥,5 ⑧,6 Item(k) Off-Diagonal Components
5.7 1.5 3.1 (Corresponding Column ID)
3
③ ⑤,7 ⑦,8 （INT, k=1, index(N)）
9.8 4.1 2.5 2.7 AMat(k) Off-Diagonal Components (Value)
4
④ ②,9 ⑤,10 ⑥,11 （ REAL, k=1, index(N) ）
11.5 3.1 9.5 10.4 4.3
5 {Y}= [A]{X}
⑤ ①,12 ②,13 ③,14 ⑦,15
12.4 6.5 9.5 do i= 1, N
6
⑥ ③,16 ⑦,17 Y(i)= D(i)*X(i)
23.1 6.4 2.5 1.4 13.1 do k= index(i-1)+1, index(i)
7
⑦ ②,18 ③,19 ⑥,20 ⑧,21 Y(i)= Y(i) + AMAT(k)*X(item(k))
51.3 9.5 1.3 9.6 3.1 enddo
8 enddo
⑧ ②,22 ③,23 ④,24 ⑥,25
FEM1D 73

• 1D-code for Static Linear-Elastic Problems by

Galerkin FEM
• Sparse Linear Solver
– Conjugate Gradient Method
– Preconditioning
• Storage of Sparse Matrices

• Program
FEM1D 74

Finite Element Procedures

• Initialization
– Control Data
– Node, Connectivity of Elements (N: Node#, NE: Elem#)
– Initialization of Arrays (Global/Element Matrices)
– Element-Global Matrix Mapping (Index, Item)
• Generation of Matrix
– Element-by-Element Operations (do icel= 1, NE)
• Element matrices
• Accumulation to global matrix
– Boundary Conditions
• Linear Solver
– Conjugate Gradient Method
FEM1D 75

Program: 1d.f (1/6)

variables and arrays
!C
!C 1D Steady-State Heat Transfer
!C FEM with Piece-wise Linear Elements
!C CG (Conjugate Gradient) Method
!C
!C d/dx(CdT/dx) + Q = 0
!C T=0@x=0
!C
program heat1D
implicit REAL*8 (A-H,O-Z)
integer :: N, NPLU, ITERmax
integer :: R, Z, P, Q, DD
real(kind=8) :: dX, RESID, EPS
real(kind=8) :: AREA, QV, COND
real(kind=8), dimension(:), allocatable :: PHI, RHS, X
real(kind=8), dimension(: ), allocatable :: DIAG, AMAT
real(kind=8), dimension(:,:), allocatable :: W
real(kind=8), dimension(2,2) :: KMAT, EMAT
integer, dimension(:), allocatable :: ICELNOD
integer, dimension(:), allocatable :: INDEX, ITEM
FEM1D 76

Variable/Arrays (1/2)
Name Type Size I/O Definition
NE I I # Element
N I O # Node
NPLU I O # Non-Zero Off-Diag. Components
IterMax I I MAX Iteration Number for CG
errno I O ERROR flag
R, Z, Q, P, I O Name of Vectors in CG
DD
dX R I Length of Each Element
Resid R O Residual for CG
Eps R I Convergence Criteria for CG
Area R I Sectional Area of Element
QV R I Heat Generation Rate/Volume/Time Q
COND R I Thermal Conductivity
FEM1D 77

Variable/Arrays (2/2)
Name Type Size I/O Definition
X R N O Location of Each Node
PHI R N O Temperature of Each Node
Rhs R N O RHS Vector
Diag R N O Diagonal Components
W R (N,4) O Work Array for CG
Amat R NPLU O Off-Diagonal Components (Value)
Index I 0:N O Number of Non-Zero Off-Diagonals at
Each ROW
Item I NPLU O Off-Diagonal Components
(Corresponding Column ID)
Icelnod I 2*NE O Node ID for Each Element
Kmat R (2,2) O Element Matrix [k]
Emat R (2,2) O Element Matrix
FEM1D 78

Program: 1d.f (2/6)

Initialization, Allocation of Arrays
!C
!C +-------+
!C | INIT. |
!C +-------+
!C===
open (11, file='input.dat', status='unknown')
read (11,*) NE
read (11,*) dX, QV, AREA, COND Control Data input.dat
read (11,*) ITERmax
read (11,*) EPS 4 NE (Number of Elements)
close (11) 1.0 1.0 1.0 1.0 x (Length of Each Elem.: L) ，Q, A, 
100 Number of MAX. Iterations for CG Solver
1.e-8 Convergence Criteria for CG Solver
N= NE + 1
allocate (PHI(N), DIAG(N), AMAT(2*N-2), RHS(N))
allocate (ICELNOD(2*NE), X(N))
allocate (INDEX(0:N), ITEM(2*N-2), W(N,4))
PHI = 0.d0 1 2 3 4
AMAT= 0.d0 1 2 3 4 5
DIAG= 0.d0
RHS= 0.d0
X= 0.d0 NE: # Element
N : # Node (NE+1)
FEM1D 79

Program: 1d.f (2/6)

Initialization, Allocation of Arrays
!C
!C +-------+
!C | INIT. |
!C +-------+
!C===
open (11, file='input.dat', status='unknown')
read (11,*) NE
read (11,*) dX, QV, AREA, COND
read (11,*) ITERmax
read (11,*) EPS
close (11)

N= NE + 1
allocate (PHI(N), DIAG(N), AMAT(2*N-2), RHS(N))
allocate (ICELNOD(2*NE), X(N))
allocate (INDEX(0:N), ITEM(2*N-2), W(N,4))
PHI = 0.d0
AMAT= 0.d0 Amat: Non-Zero Off-Diag. Comp.
DIAG= 0.d0
RHS= 0.d0 Item: Corresponding Column ID
X= 0.d0
FEM1D 80

Element/Global Operations
1 2 3 4
1 2 3 4 5 Global Node ID

Local Node ID
e e e e
1 2 1 2 1 2 1 2
 d N  d N  
T
 d N  d N  
T
 d N  d N  
T
 d N  d N  
T
  E  dV      E  dV      E  dV      E  dV   
V  dx dx  V  dx dx  V  dx dx  V  dx dx 
   N  dS   X N  dV  0    N  dS   X N  dV  0    N  dS   X N  dV  0    N  dS   X N  dV  0
T T T T T T T T

S V S V S V S V

[k ]( 3) {}( 3)  { f }(3) [k ]( 4 ) {}( 4)  { f }( 4 )

[k ](1) {}(1)  { f }(1) [k ]( 2 ) {}( 2 )  { f }( 2 )
k11( 3) k12( 3)  1( 3)   f1( 3)  k11( 4 ) k12( 4 )  1( 4 )   f1( 4) 
k11(1) k12(1)  1(1)   f1(1)  k11( 2 ) k12( 2 )  1( 2)   f1( 2)    (4) 
  ( 2)   ( 3) ( 3)   ( 3) 
  ( 3)   ( 4) ( 4)   ( 4) 
 (1) (1)   (1) 
  (1)   (2) ( 2)   (2)   2   f 2 
 2   f 2  k 21 k 22
 2   f 2  k 21 k 22
 2   f 2  k 21 k 22
k 21 k 22

1 2 3 4
1 2 3 4 5
[ K ]{}  {F }
 D1 AU11   1   B1  Number of non-zero off-
 AL     B 
 21 D2 AU 21   2   2  diag. components is 2 for
 AL31 D3 AU 31   3    B3 
  each node. This number
 AL41 D4 AU 41   4   B4 
    is 1 at boundary nodes).
 AL51 D5   5   B5 
FEM1D 81

Attention: In C program, node and

element ID’s start from 0.

1 2 3 4
1 2 3 4 5

0 1 2 3
0 1 2 3 4

e e e e
0 1 0 1 0 1 0 1
FEM1D 82

Program: 1d.f (2/6)

Initialization, Allocation of Array
!C
!C +-------+
!C | INIT. |
!C +-------+ icel
!C===
open (11, file='input.dat', status='unknown')
read (11,*) NE Icelnod[2*icel] Icelnod[2*icel+1]
read (11,*) dX, QV, AREA, COND =icel =icel+1
read (11,*) ITERmax
read (11,*) EPS
close (11)

N= NE + 1 Amat: Non-Zero Off-Diag. Comp.

allocate (PHI(N), DIAG(N), AMAT(2*N-2), RHS(N))
allocate (ICELNOD(2*NE), X(N)) Item: Corresponding Column ID
allocate (INDEX(0:N), ITEM(2*N-2), W(N,4))
PHI = 0.d0 Number of non-zero off-diag.
AMAT= 0.d0 components is 2 for each node. This
DIAG= 0.d0
RHS= 0.d0 number is 1 at boundary nodes).
X= 0.d0

Total Number of Non-Zero Off-Diag.

Components:
2*(N-2)+1+1= 2*N-2
FEM1D 83

Program: 1d.f (3/6)

Initialization, Allocation of Arrays (cont.)
do i= 1, N
X(i)= dfloat(i-1)*dX
enddo
X: X-coordinate
do icel= 1, NE component of each node
ICELNOD(2*icel-1)= icel
ICELNOD(2*icel )= icel + 1
enddo
KMAT(1,1)= +1.d0
KMAT(1,2)= -1.d0
KMAT(2,1)= -1.d0
KMAT(2,2)= +1.d0
FEM1D 84

Program: 1d.f (3/6)

Initialization, Allocation of Arrays (cont.)
do i= 1, N
X(i)= dfloat(i-1)*dX
enddo
icel
do icel= 1, NE
ICELNOD(2*icel-1)= icel
ICELNOD(2*icel )= icel + 1 Icelnod(2*icel-1) Icelnod(2*icel)
enddo =icel =icel+1
KMAT(1,1)= +1.d0
KMAT(1,2)= -1.d0
KMAT(2,1)= -1.d0
KMAT(2,2)= +1.d0
FEM1D 85

Program: 1d.f (3/6)

Initialization, Allocation of Arrays (cont.)
do i= 1, N
X(i)= dfloat(i-1)*dX
enddo
do icel= 1, NE
ICELNOD(2*icel-1)= icel  d N T d N    A  1  1
ICELNOD(2*icel )= icel + 1
enddo k (e)     dV 
   1  1
KMAT(1,1)= +1.d0
V 
dx dx  L  
KMAT(1,2)= -1.d0
KMAT(2,1)= -1.d0 [Kmat]
KMAT(2,2)= +1.d0
FEM1D 86

Program: 1d.f (4/6)

Global Matrix: Column ID for Non-Zero Off-Diag’s
!C
!C +--------------+ Number of non-zero off-diag. components is 2 for each
!C | CONNECTIVITY | node. This number is 1 at boundary nodes).
!C +--------------+
!C=== Total Number of Non-Zero Off-Diag. Components:
INDEX = 2
2*(N-2)+1+1= 2*N-2= NPLU= Index[N]
INDEX(0)= 0
# Non-Zero
INDEX(1)= 1 index(0)= 0
INDEX(N)= 1 Off-Diag.
1.1 2.4 3.2
1 2 index(1)= 2
do i= 1, N ① ② ⑤
INDEX(i)= INDEX(i) + INDEX(i-1) 3.6 4.3 2.5 3.7 9.1
enddo 2 4 index(2)= 6
② ① ④ ⑥ ⑧
NPLU= INDEX(N) 5.7 1.5 3.1
3 2 index(3)= 8
③ ⑤ ⑦
do i= 1, N 4 9.8 4.1 2.5 2.7
3 index(4)= 11
jS= INDEX(i-1) ④ ② ⑤ ⑥
if (i.eq.1) then 11.5 3.1 9.5 10.4 4.3
ITEM(jS+1)= i+1 5
⑤ ① ②
4 index(5)= 15
③ ⑦
else if
& (i.eq.N) then 12.4 6.5 9.5
6 2 index(6)= 17
ITEM(jS+1)= i-1 ⑥ ③ ⑦
else 23.1 6.4 2.5 1.4 13.1 4
ITEM(jS+1)= i-1 7 index(7)= 21
⑦ ② ③ ⑥ ⑧
ITEM(jS+2)= i+1
endif 8 51.3 9.5 1.3 9.6 3.1 4 index(8)= 25
enddo ⑧ ② ③ ④ ⑥
!C=== index(i-1)+1th~index(i) th
Non-Zero Off-Diag. Components corresponding to i-th row
FEM1D 87

Program: 1d.f (4/6)

Global Matrix: Column ID for Non-Zero Off-Diag’s
!C
!C +--------------+
!C | CONNECTIVITY | i-1 i
!C +--------------+ i-1 i i+1
!C===
INDEX = 2
INDEX(0)= 0
# Non-Zero
INDEX(1)= 1 index(0)= 0
INDEX(N)= 1 Off-Diag.
1.1 2.4 3.2
1 2 index(1)= 2
do i= 1, N ① ② ⑤
INDEX(i)= INDEX(i) + INDEX(i-1) 3.6 4.3 2.5 3.7 9.1
enddo 2 4 index(2)= 6
② ① ④ ⑥ ⑧
NPLU= INDEX(N) 5.7 1.5 3.1
3 2 index(3)= 8
③ ⑤ ⑦
do i= 1, N 4 9.8 4.1 2.5 2.7
3 index(4)= 11
jS= INDEX(i-1) ④ ② ⑤ ⑥
if (i.eq.1) then 11.5 3.1 9.5 10.4 4.3
ITEM(jS+1)= i+1 5
⑤ ① ②
4 index(5)= 15
③ ⑦
else if
& (i.eq.N) then 12.4 6.5 9.5
6 2 index(6)= 17
ITEM(jS+1)= i-1 ⑥ ③ ⑦
else 23.1 6.4 2.5 1.4 13.1 4
ITEM(jS+1)= i-1 7 index(7)= 21
⑦ ② ③ ⑥ ⑧
ITEM(jS+2)= i+1
endif 8 51.3 9.5 1.3 9.6 3.1 4 index(8)= 25
enddo ⑧ ② ③ ④ ⑥
!C=== index(i-1)+1th~index(i) th
Non-Zero Off-Diag. Components corresponding to i-th row
FEM1D 88

Program: 1d.f (5/6)

Element Matrix ~ Global Matrix
!C +-----------------+
!C | MATRIX ASSEMBLE |
!C +-----------------+
!C===
do icel= 1, NE
in1= ICELNOD(2*icel-1)
in2= ICELNOD(2*icel ) icel
X1 = X(in1) in1 in2
X2 = X(in2)
DL = dabs(X2-X1)
cK= AREA*COND/DL
EMAT(1,1)= Ck*KMAT(1,1)
EMAT(1,2)= Ck*KMAT(1,2)
EMAT(2,1)= Ck*KMAT(2,1)
EMAT(2,2)= Ck*KMAT(2,2)
DIAG(in1)= DIAG(in1) + EMAT(1,1)
DIAG(in2)= DIAG(in2) + EMAT(2,2)
if (icel.eq.1) then
k1= INDEX(in1-1) + 1
else
k1= INDEX(in1-1) + 2
endif
k2= INDEX(in2-1) + 1
AMAT(k1)= AMAT(k1) + EMAT(1,2)
AMAT(k2)= AMAT(k2) + EMAT(2,1)
QN= 0.50d0*QV*AREA*DL
RHS(in1)= RHS(in1) + QN
RHS(in2)= RHS(in2) + QN
enddo
!C===
FEM1D 89

Program: 1d.f (5/6)

Element Matrix ~ Global Matrix
!C +-----------------+
!C | MATRIX ASSEMBLE |
!C +-----------------+
!C===
do icel= 1, NE
in1= ICELNOD(2*icel-1)
in2= ICELNOD(2*icel ) icel
X1 = X(in1) in1 in2
X2 = X(in2)
DL = dabs(X2-X1)
cK= AREA*COND/DL
 1  1 A
Emat   k (e)  A 
EMAT(1,1)= Ck*KMAT(1,1)
EMAT(1,2)= Ck*KMAT(1,2)
EMAT(2,1)= Ck*KMAT(2,1)   Kmat 
EMAT(2,2)= Ck*KMAT(2,2) L  1  1 L
DIAG(in1)= DIAG(in1) + EMAT(1,1)
DIAG(in2)= DIAG(in2) + EMAT(2,2)
if (icel.eq.1) then
k1= INDEX(in1-1) + 1
else
k1= INDEX(in1-1) + 2
endif
k2= INDEX(in2-1) + 1
AMAT(k1)= AMAT(k1) + EMAT(1,2)
AMAT(k2)= AMAT(k2) + EMAT(2,1)
QN= 0.50d0*QV*AREA*DL
RHS(in1)= RHS(in1) + QN
RHS(in2)= RHS(in2) + QN
enddo
!C===
FEM1D 90

Program: 1d.f (5/6)

Element Matrix ~ Global Matrix
!C +-----------------+
!C | MATRIX ASSEMBLE |
!C +-----------------+
!C===
do icel= 1, NE
in1= ICELNOD(2*icel-1)
in2= ICELNOD(2*icel ) icel
X1 = X(in1) in1 in2
X2 = X(in2)
DL = dabs(X2-X1)
cK= AREA*COND/DL
EMAT(1,1)= Ck*KMAT(1,1)
EMAT(1,2)= Ck*KMAT(1,2)
EMAT(2,1)= Ck*KMAT(2,1)
EMAT(2,2)= Ck*KMAT(2,2)
EA  1  1
DIAG(in1)= DIAG(in1) + EMAT(1,1)
Emat   k 
(e)

L   1  1
DIAG(in2)= DIAG(in2) + EMAT(2,2)
if (icel.eq.1) then
k1= INDEX(in1-1) + 1
else
k1= INDEX(in1-1) + 2
endif
k2= INDEX(in2-1) + 1
AMAT(k1)= AMAT(k1) + EMAT(1,2)
AMAT(k2)= AMAT(k2) + EMAT(2,1)
QN= 0.50d0*QV*AREA*DL
RHS(in1)= RHS(in1) + QN
RHS(in2)= RHS(in2) + QN
enddo
!C===
FEM1D 91

Program: 1d.f (5/6)

Element Matrix ~ Global Matrix
!C +-----------------+
!C | MATRIX ASSEMBLE |
!C +-----------------+
!C===
do icel= 1, NE
in1= ICELNOD(2*icel-1)
in2= ICELNOD(2*icel ) icel
X1 = X(in1) in1 in2
X2 = X(in2)
DL = dabs(X2-X1)
cK= AREA*COND/DL
EMAT(1,1)= Ck*KMAT(1,1)
EMAT(1,2)= Ck*KMAT(1,2)
EMAT(2,1)= Ck*KMAT(2,1)
EMAT(2,2)= Ck*KMAT(2,2)
DIAG(in1)= DIAG(in1) + EMAT(1,1) Non-zero Off-Diag. at i-th row:
DIAG(in2)= DIAG(in2) + EMAT(2,2) Index(i-1)+1, Index(i-1)+2
if (icel.eq.1) then
k1= INDEX(in1-1) + 1
else i-1 i i+1
k1= INDEX(in1-1) + 2
endif INDEX(i-1)+1 INDEX(i-1)+2
k2= INDEX(in2-1) + 1
AMAT(k1)= AMAT(k1) + EMAT(1,2)
A  1  1 k1
AMAT(k2)= AMAT(k2) + EMAT(2,1)
Emat   k (e)

QN= 0.50d0*QV*AREA*DL L   1  1
RHS(in1)= RHS(in1) + QN
RHS(in2)= RHS(in2) + QN
enddo k2
!C===
FEM1D 92

General Elements: k1
“in2” as a off-diag. component of “in1”
!C +-----------------+
!C | MATRIX ASSEMBLE |
!C +-----------------+
!C===
do icel= 1, NE
in1= ICELNOD(2*icel-1)
in2= ICELNOD(2*icel )
X1 = X(in1)
X2 = X(in2)
DL = dabs(X2-X1) icel
in1 in2
cK= AREA*COND/DL
EMAT(1,1)= Ck*KMAT(1,1)
EMAT(1,2)= Ck*KMAT(1,2)
EMAT(2,1)= Ck*KMAT(2,1)
EMAT(2,2)= Ck*KMAT(2,2)
DIAG(in1)= DIAG(in1) + EMAT(1,1) Non-zero Off-Diag. at i-th row:
DIAG(in2)= DIAG(in2) + EMAT(2,2) Index(i-1)+1, Index(i-1)+2
if (icel.eq.1) then
k1= INDEX(in1-1) + 1
else i-1 in1 in2
k1= INDEX(in1-1) + 2
endif INDEX(i-1)+1 INDEX(i-1)+2
k2= INDEX(in2-1) + 1
AMAT(k1)= AMAT(k1) + EMAT(1,2)
A  1  1 k1
AMAT(k2)= AMAT(k2) + EMAT(2,1)
Emat   k (e)

QN= 0.50d0*QV*AREA*DL L   1  1
RHS(in1)= RHS(in1) + QN
RHS(in2)= RHS(in2) + QN
enddo
!C===
FEM1D 93

General Elements: k2
“in1” as a off-diag. component of “in2”
!C +-----------------+
!C | MATRIX ASSEMBLE |
!C +-----------------+
!C===
do icel= 1, NE
in1= ICELNOD(2*icel-1)
in2= ICELNOD(2*icel )
X1 = X(in1)
X2 = X(in2)
DL = dabs(X2-X1) icel
in1 in2
cK= AREA*COND/DL
EMAT(1,1)= Ck*KMAT(1,1)
EMAT(1,2)= Ck*KMAT(1,2)
EMAT(2,1)= Ck*KMAT(2,1)
EMAT(2,2)= Ck*KMAT(2,2)
DIAG(in1)= DIAG(in1) + EMAT(1,1) Non-zero Off-Diag. at i-th row:
DIAG(in2)= DIAG(in2) + EMAT(2,2) Index(i-1)+1, Index(i-1)+2
if (icel.eq.1) then
k1= INDEX(in1-1) + 1
else in1 in2 i+1
k1= INDEX(in1-1) + 2
endif INDEX(i-1)+1 INDEX(i-1)+2
k2= INDEX(in2-1) + 1
AMAT(k1)= AMAT(k1) + EMAT(1,2)
A  1  1
AMAT(k2)= AMAT(k2) + EMAT(2,1)
Emat   k (e)

QN= 0.50d0*QV*AREA*DL L   1  1
RHS(in1)= RHS(in1) + QN
RHS(in2)= RHS(in2) + QN
enddo k2
!C===
FEM1D 94

0-th Element: k1
“in2” as a off-diag. component of “in1”
!C +-----------------+
!C | MATRIX ASSEMBLE |
!C +-----------------+
!C===
do icel= 1, NE
in1= ICELNOD(2*icel-1)
in2= ICELNOD(2*icel )
X1 = X(in1)
X2 = X(in2)
DL = dabs(X2-X1) Icel＝1
in1 in2
cK= AREA*COND/DL
EMAT(1,1)= Ck*KMAT(1,1)
EMAT(1,2)= Ck*KMAT(1,2)
EMAT(2,1)= Ck*KMAT(2,1)
EMAT(2,2)= Ck*KMAT(2,2)
DIAG(in1)= DIAG(in1) + EMAT(1,1)
Non-zero Off-Diag. at i-th row:
DIAG(in2)= DIAG(in2) + EMAT(2,2) Index(i-1)+1 only
if (icel.eq.1) then
k1= INDEX(in1-1) + 1
else in1 in2
k1= INDEX(in1-1) + 2
endif INDEX(i-1)+1
k2= INDEX(in2-1) + 1
AMAT(k1)= AMAT(k1) + EMAT(1,2)
A  1  1 k1
AMAT(k2)= AMAT(k2) + EMAT(2,1)
Emat   k 
(e)

QN= 0.50d0*QV*AREA*DL L   1  1
RHS(in1)= RHS(in1) + QN
RHS(in2)= RHS(in2) + QN
enddo
!C===
FEM1D 95

Program: 1d.f (5/6)

RHS: Heat Generation Term
!C +-----------------+
!C | MATRIX ASSEMBLE |
!C +-----------------+
!C===
do icel= 1, NE
in1= ICELNOD(2*icel-1)
in2= ICELNOD(2*icel ) icel
X1 = X(in1) in1 in2
X2 = X(in2)
DL = dabs(X2-X1)
cK= AREA*COND/DL
EMAT(1,1)= Ck*KMAT(1,1)
EMAT(1,2)= Ck*KMAT(1,2)
EMAT(2,1)= Ck*KMAT(2,1)
EMAT(2,2)= Ck*KMAT(2,2)
DIAG(in1)= DIAG(in1) + EMAT(1,1)
DIAG(in2)= DIAG(in2) + EMAT(2,2)
if (icel.eq.1) then
k1= INDEX(in1-1) + 1
else
k1= INDEX(in1-1) + 2
endif
k2= INDEX(in2-1) + 1
AMAT(k1)= AMAT(k1) + EMAT(1,2)
AMAT(k2)= AMAT(k2) + EMAT(2,1)
L
1  x / L  Q AL 1
V QN  dV  QA0  x / L  dx  2 1
QN= 0.50d0*QV*AREA*DL
RHS(in1)= RHS(in1) + QN  T 
RHS(in2)= RHS(in2) + QN
enddo
!C===
FEM1D 96

Program: 1d.f (6/6)

Dirichlet B.C. @ X=0
!C
!C +---------------------+
!C | BOUNDARY CONDITIONS |
!C +---------------------+
!C===

!C
!C-- X=Xmin
i= 1
jS= INDEX(i-1)

AMAT(jS+1)= 0.d0
DIAG(i)= 1.d0
RHS (i)= 0.d0

do k= 1, NPLU
if (ITEM(k).eq.1) AMAT(k)= 0.d0
enddo
!C===
FEM1D 97

1D Steady State Heat Conduction

heat generation/volume Q
  T  
 Q  0
x  x 

x=0 (xmin) x= xmax

• Uniform: Sectional Area: A, Thermal Conductivity: 

• Heat Generation Rate/Volume/Time [QL-3T-1] Q
• Boundary Conditions
– x=0 ： T= 0 (Fixed Temperature)
T
– x=xmax：  0 (Insulated)
x
FEM1D 98

(Linear) Equation at x=0

T1= 0 (or T0 = 0)
heat generation/volume Q
  T  
 Q  0
x  x 

x=0 (xmin) x= xmax

• Uniform: Sectional Area: A, Thermal Conductivity: 

• Heat Generation Rate/Volume/Time [QL-3T-1] Q
• Boundary Conditions
– x=0 ： T= 0 (Fixed Temperature)
T
– x=xmax：  0 (Insulated)
x
FEM1D 99

Program: 1d.f (6/6)

Dirichlet B.C. @ X=0
!C
!C +---------------------+
!C | BOUNDARY CONDITIONS | T1=0
!C +---------------------+
!C===
Diagonal Component=1
!C RHS=0
!C-- X=Xmin
i= 1
Off-Diagonal Components= 0.
jS= INDEX(i-1)

AMAT(jS+1)= 0.d0
DIAG(i)= 1.d0
RHS (i)= 0.d0

do k= 1, NPLU
if (ITEM(k).eq.1) AMAT(k)= 0.d0
enddo
!C===
FEM1D 100

Program: 1d.f (6/6)

Dirichlet B.C. @ X=0
!C
!C +---------------------+
!C | BOUNDARY CONDITIONS | T1=0
!C +---------------------+
!C===
Diagonal Component=1
!C RHS=0
!C-- X=Xmin
i= 1
Off-Diagonal Components= 0.
jS= INDEX(i-1)
Erase !
AMAT(jS+1)= 0.d0
DIAG(i)= 1.d0
RHS (i)= 0.d0

do k= 1, NPLU
if (ITEM(k).eq.1) AMAT(k)= 0.d0
enddo
!C===
FEM1D 101

Program: 1d.f (6/6)

Dirichlet B.C. @ X=0
!C
!C +---------------------+ T1=0
!C | BOUNDARY CONDITIONS |
!C +---------------------+ Diagonal Component=1
!C===
RHS=0
!C Off-Diagonal Components= 0.
!C-- X=Xmin
i= 1
Elimination and Erase
jS= INDEX(i-1)

AMAT(jS+1)= 0.d0
DIAG(i)= 1.d0
RHS (i)= 0.d0

do k= 1, NPLU
if (ITEM(k).eq.1) AMAT(k)= 0.d0
enddo
!C===

Column components of boundary nodes (Dirichlet B.C.) are

moved to RHS and eliminated for keeping symmetrical feature of
the matrix (in this case just erase off-diagonal components)
FEM1D 102

if T1≠ 0
!C
!C +---------------------+
!C | BOUNDARY CONDITIONS |
Column components of boundary nodes
!C +---------------------+ (Dirichlet B.C.) are moved to RHS and
!C=== eliminated for keeping symmetrical feature
!C of the matrix.
!C-- X=Xmin
i= 1
Index[ j 1]1

 Amat
jS= INDEX(i-1)
Diag j j  k Item[ k ]  Rhs j
AMAT(jS+1)= 0.d0 k  Index[ j ]
DIAG(i)= 1.d0
RHS (i)= PHImin

do i= 1, N
do k= INDEX(i-1)+1, INDEX(i)
if (ITEM(k).eq.1) then
RHS (i)= RHS(i) – AMAT(k)*PHImin
AMAT(k)= 0.d0
endif
enddo
enddo
!C===
FEM1D 103

if T1≠ 0
!C
!C +---------------------+ Index[ j 1]1
!C | BOUNDARY CONDITIONS |
!C +---------------------+
!C===
Diag j j   Amat
k  Index[ j ], k  k s
k Item[ k ]

!C  Rhs j  Amatk s Item[ k s ]

!C-- X=Xmin
i= 1
jS= INDEX(i-1)
 Rhs j  Amatk s min where Item[k s ]  0
AMAT(jS+1)= 0.d0
DIAG(i)= 1.d0
RHS (i)= PHImin

do i= 1, N
do k= INDEX(i-1)+1, INDEX(i)
if (ITEM(k).eq.1) then
RHS (i)= RHS(i) – AMAT(k)*PHImin
AMAT(k)= 0.d0
endif Column components of boundary nodes
enddo
enddo
(Dirichlet B.C.) are moved to RHS and
!C=== eliminated for keeping symmetrical feature
of the matrix.
FEM1D 104

Secondary B.C. (Insulated)

Heat Gen. Rate Q
  T  
 Q  0
x  x 

x=0 (xmin) x= xmax

T  0@ x  0 T
 0 @ x  xmax
x

0
 q N  dS  q A
dT
 q A , q  
T
xL Surface Flux
S 1 dx

According to insulated B.C., q  0

is satisfied. No contribution by this term.
T Insulated B.C. is automatically satisfied
 0 @ x  xmax without explicit operations
x -> Natural B.C.
FEM1D 105

Preconditioned CG Solver
Compute r(0)= b-[A]x(0)
for i= 1, 2, …
D1 0 ... 0 0
0 D 0 0 
solve [M]z(i-1)= r(i-1)
 2 
i-1= r(i-1) z(i-1)
if i=1
M    ... ... ... 
 
p(1)= z(0)  0 0 DN 1 0 
else  0 0 ... 0 DN 
i-1= i-1/i-2
p(i)= z(i-1) + i-1 p(i-1)
endif
q(i)= [A]p(i)
i = i-1/p(i)q(i)
x(i)= x(i-1) + ip(i)
r(i)= r(i-1) - iq(i)
check convergence |r|
end
FEM1D 106

Diagonal Scaling, Point-Jacobi

D1 0 ... 0 0
0 D 0 0 
 2 
M    ... ... ... 
 
0 0 DN 1 0 
 0 0 ... 0 DN 

• solve [M]z(i-1)= r(i-1) is very easy.

• Provides fast convergence for simple problems.
• 1d.f, 1d.c
FEM1D 107

CG Solver (1/6)
!C
!C +---------------+
!C | CG iterations |
!C +---------------+
!C===
R = 1
Reciprocal numbers （逆数）of diagonal
Z = 2 components are stored in W(i,DD).
Q = 2 Computational cost for division is
P = 3
usually expensive.
DD= 4

do i= 1, N
W(i,DD)= 1.0D0 / DIAG(i)
enddo
FEM1D 108

CG Solver (1/6)
!C
!C +---------------+ Compute r(0)= b-[A]x(0)
!C | CG iterations | for i= 1, 2, …
!C +---------------+ solve [M]z(i-1)= r(i-1)
!C===
R = 1
i-1= r(i-1) z(i-1)
Z = 2 if i=1
Q = 2 p(1)= z(0)
P = 3 else
DD= 4
i-1= i-1/i-2
do i= 1, N p(i)= z(i-1) + i-1 p(i-1)
W(i,DD)= 1.0D0 / DIAG(i) endif
enddo
q(i)= [A]p(i)
i = i-1/p(i)q(i)
W(i,1)= W(i,R) ⇒ {r} x(i)= x(i-1) + ip(i)
W(i,2)= W(i,Z) ⇒ {z} r(i)= r(i-1) - iq(i)
check convergence |r|
W(i,2)= W(i,Q) ⇒ {q} end
W(i,3)= W(i,P) ⇒ {p}
W(i,4)= W(i,DD) ⇒ 1/{D}
FEM1D 109

CG Solver (2/6)
!C Compute r(0)= b-[A]x(0)
!C-- {r0}= {b} - [A]{xini} | for i= 1, 2, …
!C 初期残差 solve [M]z(i-1)= r(i-1)
do i= 1, N i-1= r(i-1) z(i-1)
W(i,R) = DIAG(i)*PHI(i) if i=1
do j= INDEX(i-1)+1, INDEX(i)
W(i,R) = W(i,R) + AMAT(j)*PHI(ITEM(j))
p(1)= z(0)
enddo else
enddo i-1= i-1/i-2
p(i)= z(i-1) + i-1 p(i-1)
BNRM2= 0.0D0
do i= 1, N endif
BNRM2 = BNRM2 + RHS(i) **2 q(i)= [A]p(i)
W(i,R)= RHS(i) - W(i,R) i = i-1/p(i)q(i)
enddo
x(i)= x(i-1) + ip(i)
BNRM2=|b|2 r(i)= r(i-1) - iq(i)
for convergence criteria check convergence |r|
of CG solvers end
FEM1D 110

CG Solver (3/6)
do iter= 1, ITERmax Compute r(0)= b-[A]x(0)
for i= 1, 2, …
!C
!C-- {z}= [Minv]{r} solve [M]z(i-1)= r(i-1)
i-1= r(i-1) z(i-1)
do i= 1, N if i=1
W(i,Z)= W(i,DD) * W(i,R)
enddo
p(1)= z(0)
else
!C i-1= i-1/i-2
!C-- RHO= {r}{z} p(i)= z(i-1) + i-1 p(i-1)
RHO= 0.d0 endif
do i= 1, N q(i)= [A]p(i)
RHO= RHO + W(i,R)*W(i,Z) i = i-1/p(i)q(i)
enddo
x(i)= x(i-1) + ip(i)
r(i)= r(i-1) - iq(i)
check convergence |r|
end
FEM1D 111

CG Solver (4/6)
!C Compute r(0)= b-[A]x(0)
!C-- {p} = {z} if ITER=1 for i= 1, 2, …
!C BETA= RHO / RHO1 otherwise
solve [M]z(i-1)= r(i-1)
if ( iter.eq.1 ) then i-1= r(i-1) z(i-1)
do i= 1, N if i=1
W(i,P)= W(i,Z)
enddo
p(1)= z(0)
else else
BETA= RHO / RHO1 i-1= i-1/i-2
do i= 1, N p(i)= z(i-1) + i-1 p(i-1)
W(i,P)= W(i,Z) + BETA*W(i,P)
enddo endif
endif q(i)= [A]p(i)
i = i-1/p(i)q(i)
!C
!C-- {q}= [A]{p}
x(i)= x(i-1) + ip(i)
r(i)= r(i-1) - iq(i)
do i= 1, N check convergence |r|
W(i,Q) = DIAG(i)*W(i,P) end
do j= INDEX(i-1)+1, INDEX(i)
W(i,Q) = W(i,Q) + AMAT(j)*W(ITEM(j),P)
enddo
enddo
FEM1D 112

CG Solver (5/6)
!C Compute r(0)= b-[A]x(0)
!C-- ALPHA= RHO / {p}{q} for i= 1, 2, …
C1= 0.d0 solve [M]z(i-1)= r(i-1)
do i= 1, N i-1= r(i-1) z(i-1)
C1= C1 + W(i,P)*W(i,Q) if i=1
enddo
ALPHA= RHO / C1
p(1)= z(0)
else
!C i-1= i-1/i-2
!C-- {x}= {x} + ALPHA*{p} p(i)= z(i-1) + i-1 p(i-1)
!C {r}= {r} - ALPHA*{q}
endif
do i= 1, N q(i)= [A]p(i)
PHI(i)= PHI(i) + ALPHA * W(i,P) i = i-1/p(i)q(i)
W(i,R)= W(i,R) - ALPHA * W(i,Q)
enddo
x(i)= x(i-1) + ip(i)
r(i)= r(i-1) - iq(i)
check convergence |r|
end
FEM1D 113

CG Solver (6/6)
DNRM2 = 0.0 Compute r(0)= b-[A]x(0)
do i= 1, N for i= 1, 2, …
DNRM2= DNRM2 + W(i,R)**2
enddo solve [M]z(i-1)= r(i-1)
i-1= r(i-1) z(i-1)
RESID= dsqrt(DNRM2/BNRM2) if i=1
if ( RESID.le.EPS) goto 900
p(1)= z(0)
RHO1 = RHO i-2 else
i-1= i-1/i-2
p(i)= z(i-1) + i-1 p(i-1)
enddo
900 continue endif
q(i)= [A]p(i)
i = i-1/p(i)q(i)
DNorm2 r Ax  b x(i)= x(i-1) + ip(i)
Resid     Eps
BNorm2 b b r(i)= r(i-1) - iq(i)
check convergence |r|
Control Data input.dat end
4 NE (Number of Elements)
1.0 1.0 1.0 1.0 x (Length of Each Elem.: L) ，Q, A, 
100 Number of MAX. Iterations for CG Solver
1.e-8 Convergence Criteria for CG Solver
FEM1D 114

Finite Element Procedures

• Initialization
– Control Data
– Node, Connectivity of Elements (N: Node#, NE: Elem#)
– Initialization of Arrays (Global/Element Matrices)
– Element-Global Matrix Mapping (Index, Item)
• Generation of Matrix
– Element-by-Element Operations (do icel= 1, NE)
• Element matrices
• Accumulation to global matrix
– Boundary Conditions
• Linear Solver
– Conjugate Gradient Method
FEM1D 115

Remedies for Higher Accuracy

• Finer Meshes
log A1 x  A2   log A2 
NE=8, dX=12.5 F
 u
8 iters, RESID= 2.822910E-16 U(N)= 1.953586E-01 EA1
### DISPLACEMENT
1 0.000000E+00 -0.000000E+00
2 1.101928E-02 1.103160E-02
3 2.348034E-02 2.351048E-02
4 3.781726E-02 3.787457E-02
5 5.469490E-02 5.479659E-02 1 2 3 4 5
6 7.520772E-02 7.538926E-02
7 1.013515E-01 1.016991E-01 1 2 3 4
8 1.373875E-01 1.381746E-01
9 1.953586E-01 1.980421E-01

NE=20, dX=5

20 iters, RESID= 5.707508E-15 U(N)= 1.975734E-01

### DISPLACEMENT
1 0.000000E+00 -0.000000E+00
2 4.259851E-03 4.260561E-03
3 8.719160E-03 8.720685E-03
4 1.339752E-02 1.339999E-02
……
17 1.145876E-01 1.146641E-01
18 1.295689E-01 1.296764E-01
19 1.473466E-01 1.475060E-01
20 1.692046E-01 1.694607E-01
21 1.975734E-01 1.980421E-01
FEM1D 116

Remedies for Higher Accuracy

• Finer Meshes
• Higher Order Shape/Interpolation Function（高次補
間関数・形状関数）
– Higher-Order Element（高次要素）
– Linear-Element, 1st-Order Element: Lower Order（低次要
素）
• Formulation which assures continuity of n-th order
derivatives
– Cn Continuity（Cn連続性）
FEM1D 117

Remedies for Higher Accuracy

• Finer Meshes
• Higher Order Shape/Interpolation Function（高次補
間関数・形状関数）
– Higher-Order Element（高次要素）
– Linear-Element, 1st-Order Element: Lower Order（低次要
素）
• Formulation which assures continuity of n-th order
derivatives
– Cn Continuity（Cn連続性）
• Linear Elements
– Piecewise Linear
– C0 Continuity
• Only dependent variables are continuous at element boundary
FEM1D 118

Example: 1D Heat Transfer (1/2)

T0  0C , h  0.10W / cm 2  K • Temp. Thermal Fins
• Circular Sectional Area,
  4.00W / cm  K 銅Cupper r=1cm

• Boundary Condition
7.50cm
– x=0 : Fixed Temperature
TS  150C – x=7.5 : Insulated
• Convective Heat Transfer
on Cylindrical Surface
– q= h (T-T0)
– q：Heat Flux
Convective Heat Transfer on
Cylindrical Surface • Heat Flow/Unit Surface
Area/sec.
FEM1D 119

Example: 1D Heat Transfer (2/2)

150
### RESULTS (linear interpolation)
ID X FEM. ANALYTICAL Exact Solution
1 0.00000 150.00000 150.00000 ERR(%): 0.00000
2 1.87500 102.62226 103.00165 ERR(%): 0.25292
3 3.75000 73.82803 74.37583 ERR(%): 0.36520 100

deg-C
4 5.62500 58.40306 59.01653 ERR(%): 0.40898
5 7.50000 53.55410 54.18409 ERR(%): 0.41999

### RESULTS (quadratic interpolation) 50  hP 

ID X FEM. ANALYTICAL cosh  x 
1 0.00000 150.00000 150.00000 ERR(%): 0.00000 T ( x)  TS  T0   A  T
0
 hP 
2 1.87500 102.98743 103.00165 ERR(%): 0.00948 cosh X max 
3 3.75000 74.40203 74.37583 ERR(%): 0.01747  A 
4 5.62500 59.02737 59.01653 ERR(%): 0.00722 0
5 7.50000 54.21426 54.18409 ERR(%): 0.02011 0.00 2.00 4.00 6.00 8.00

### RESULTS (linear interpolation) X

ID X FEM. ANALYTICAL
1 0.00000 150.00000 150.00000 ERR(%): 0.00000
2 0.93750 123.71561 123.77127 ERR(%): 0.03711 Quadratic interpolation provides
3 1.87500 102.90805 103.00165 ERR(%): 0.06240 more accurate solution,
4 2.81250 86.65618 86.77507 ERR(%): 0.07926
5 3.75000 74.24055 74.37583 ERR(%): 0.09019 especially if X is close to 7.50cm.
6 4.68750 65.11151 65.25705 ERR(%): 0.09703
7 5.62500 58.86492 59.01653 ERR(%): 0.10107
8 6.56250 55.22426 55.37903 ERR(%): 0.10317
9 7.50000 54.02836 54.18409 ERR(%): 0.10382