Scribd
Upload
26 views

y A A x: ← y+Ax ← A+xy ∈ R ,y,z ∈ R ← y + Ax ← A + xy ,Φ ← A + Bc

This document discusses computer arithmetic and floating point operations. It covers: 1. The IEEE floating point standard and its representation of numbers using a sign bit, exponent field, and significand field. It also discusses rounding and precision limits. 2. Performance characteristics of basic linear algebra subprograms (BLAS) like vector-vector addition (MV) and generalized rank-1 update (GER). FLOP and word counts are provided. 3. Limits of real data types like realmax, realmin, and the difference between the maximum representable number and the next larger number. 4. The floating point dot product operation and how the fused multiply-add (FMA) instruction can improve its

Uploaded by

Kostas Koutsoukos
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
26 views

y A A x: ← y+Ax ← A+xy ∈ R ,y,z ∈ R ← y + Ax ← A + xy ,Φ ← A + Bc

This document discusses computer arithmetic and floating point operations. It covers: 1. The IEEE floating point standard and its representation of numbers using a sign bit, exponent field, and significand field. It also discusses rounding and precision limits. 2. Performance characteristics of basic linear algebra subprograms (BLAS) like vector-vector addition (MV) and generalized rank-1 update (GER). FLOP and word counts are provided. 3. Limits of real data types like realmax, realmin, and the difference between the maximum representable number and the next larger number. 4. The floating point dot product operation and how the fused multiply-add (FMA) instruction can improve its

Uploaded by

Kostas Koutsoukos
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

 


            I



I.

  

     


        
  
 
    
         !"  
#   $



   %      
   &
'        


   

 (
        
     ) "  
 

 
& 


 
         
* " +     
& 

&
  # * " 

&    
#  
&  load
& &  


&  store* " ,
 
   
   
& 
  
 
#  
   
*
" -   

 
   
&  &      #
* " 
&  miss.
 
#
    
&   #  !  
/   # 0 "
1 $   &

    , min    
 y y + Ax  A A + xy !  A Rnn  x, y, z Rn "
 
 

&     & BLAS-2

 2   MV y y + Ax) = n(2n 1), min = n2 + 3n 2   GER A A + xy ) = 2n2 , min =
2n2 + 2n 3   

.  
  
 
 
# 
& 
       n
4 &.     #     A A + BC 
        5    

    #
&  2 '         BLAS-2
II.

 6      
   IEEE   &
 !  1.      71 bits 
 

  /0
   #&. 
&     &.
&  "   & 

     
  
 
     
/  0   /  
8
    
0  4 
  
&      
 u9

 :    
  
     
  #     
 

&     .       
&
& x = 1.  .  
  
 t = 53

 
    
   & 
   
 !  &;

 " 4    
  
 
&
u

= max
x

|x fl(x)| |0.0 0  |
=
|x|
1.0

= 2t + 2t1 + = 2t (1 + 21 + ) = 2t

1
= 2t+1 = 252 .
(1 1/2)

! 2   
        
. &    /max0.      


   #   /sup0"
1 6   
 realmax, realmin
 &     
     
 +



  
    
)
i) realmax+ realmin ii) realmin/0 iii) realmin/2 == 0 iv) realmax+realmax/2


 i) realmax+ realmin = realmax     & 

 


&  

realmax+ realmin = realmax (1+realmin/realmax)   
 
& 
 #
  
 realmin/realmax < eps   eps
&      (   
 . 

&  



  realmin
& 8 
&
 
    MATLAB 8
 
 
       .
& < !  realmin = -realmax ii) realmin/0 = 
   
  .   
 
&  Inf = .   

.

 
<.    
 

& -Inf iii) realmin/2 == 0 4#     

  
&  .  

  &  
& .   
   realmin/2

 
& <.
    
&   < /0 iv) realmax+realmax/2 :   
& 

   
&
     
  
 #
Inf > 
    
#
    
  (i) &
    & 


 
&;
   
realmax/2
5 $   &

 #    
    #  
   
    &
s=0; for j = 1:4, s = s+ x(j)y(j); end;

      DOT            &    
&
 FMA
!? 
)  FMA
 
    #   
 
"

  
 
 
 s  fl(s)
&          
. & 
6  
 FMA 
& 
) s = FMA(t,a,b) 
 #
s = t+a*b +
   
& 

& 
)
s=0; for j = 1:4, s = FMA(s,x(j),y(j)); end; + # 
& 
)
fl(s) = ((((s + x(1)y(1))(1 + 1 ) + x(2)y(2))(1 + 2 ) + x(3)y(3))(1 + 3 ) + x(4)y(4))(1 + 4 )
= x(1)y(1)(1 + 4 ) + x(2)y(2)(1 + 3 ) + x(3)y(3)(1 + 2 ) + x(4)y(4)(1 + 1 )

 .  . | j | u  | j | j = ju/(1 ju). 4     # # 


 
)
|fl(s) s| = |x(1)y(1)4 + x(2)y(2)3 + x(3)y(3)2 + x(4)y(4)1 |
4

4 |x( j)y( j)|.


j=1

> 


&    
|fl(s) s| |s|4 .

III. @
     
    
     &
#    
  
 
#      
   

&
   
     )
s=x(1)*x(1); for j = 2:n, s = s+ x(j)*x(j); end;

  
 
& x(j)
&          x( j) x( j) < Inf
 $   &


&       
&   & # 
  

 =    
.
   


&  +  
 s
&      
x) f (x) = nj=1 2j .    & #
j   x( j) 4 . f : Rn R+ .   R+ &;

       4 .  A   f 
&   ! " JR1n )
J

f
f
,...,
]
1
n
= [21 , ..., 2n ] = 2x.
= [

-
&    
 
&   &;
 
)
cond( f ; x) = J

x
x2
x
= 2x
=2
s
|s|
s

#  s
&      4 . 



&
  .    
     s = x22  
&  &
 cond( f ; x) = 2
2   & #.     
)
fl(s) = 21 < n > +22 < n > +23 < n 1 > + 2n < 2 >
j
 
 < j >= k=1
(1 + k ) 4    
  
&  
 
&       

     
&  

x := [1 (1 + n )1/2 , 2 (1 + n )1/2 , 3 (1 + n1 )1/2 , ..., n (1 + 2 )1/2 ]

+  
&  x
&   &   x 4 
.
x x = [1 (1 + n )1/2 1 , 2 (1 + n )1/2 2 , 3 (1 + n1 )1/2 3 , ..., n (1 + 2 )1/2 n ]
n x

1 $   &

 #    
  
  #      
 s

 =   
 &.  

|fl(s) s|
2n .
|s|

IV.

 $     

     
        
& /
   0 

7 -    
 

&   
     !   " 
  n
s = 0.0; for j=1:n, s = s+x(j)*y(j); end


 @ 
 
.    MATLAB .     rem(x, y)
 #
    
&
    x, y -   
& 
)
m = rem(n,5); s = 0.0;
for j=1:m, s = s+x(j)*y(j); end
for j=m+1:5:n
s = s+x(j)*y(j)+x(j+1)*y(j+1)+x(j+2)*y(j+2)+x(j+3)*y(j+3)+x(j+4)*y(j+4);
end

1 @
   &
 y y + Ak x.   k > 1
& 
    . A Rnn  x, y Rn  $


   &     &

 , min   & ! .   &  

 
    
    
  kn3 "

 6     #
&   &
& )
LOAD A, x, y
for kstep = 1 : k
for i = 1 : n
z(i) = 0
for j = 1 : n
z(i) = z(i) + A(i, j)x( j)
end
end
x=z
end
y = y + x;
STORE y

+   
& = 2kn2 + n  min = n2 + 3n > 
    &
 
)
y = y + A(A( A(Ax) )     
 
 .
&   #    &;
 

&  ! = 2n2 " &  


&   ! = 2n3 " , .     
 
& 
 #                .  

   8


. 
&
 #) y = y + ( ((A2 )(A2 )) (A2 )(A2 )) )    ;
O(n3 log2 k) 


&    

 

 

  )
LOAD A, x, y
for kstep = 1 : k
for i = 1 : n
z(i) = A(i, 1) x(1)
for j = 2 : n
z(i) = z(i) + A(i, j)x( j)
end
end
x=z
end
y = y + x;
STORE y

 
= kn(2n 1) + n
5 

   

.    &

   & 
 
 # 

 K = O(n)

 
&      

  ! #  
&  min " ,.


&  

= min 
 #  O(n). 

&      A

 

 

  
LOAD x
for kstep = 1 : k
for i = 1 : n
z(i) = 0
for j = 1 : n
LOAD A(i, j)
z(i) = z(i) + A(i, j)x( j)
end
end
x=z
end
for i = 1 : n
y(i) = y(i) + x(i);
end
STORE y

+   .  
&
 
 #  
 3n !   x, y, z" (
      .
= 3n + kn2 

You might also like