Skip to content

Instantly share code, notes, and snippets.

@weslleyspereira
Last active September 1, 2021 19:23
Show Gist options
  • Star 0 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
  • Save weslleyspereira/8144808440c2523b871a51f15a7932f9 to your computer and use it in GitHub Desktop.
Save weslleyspereira/8144808440c2523b871a51f15a7932f9 to your computer and use it in GitHub Desktop.
Test double complex division in Fortran
*> \brief zabs tests the robustness and precision of the intrinsic ABS for double complex
*> \author Weslley S. Pereira, University of Colorado Denver, U.S.
*
*> \verbatim
*>
*> Real values for test:
*> (1) x = 2**m, where m = MINEXPONENT-DIGITS, ..., MINEXPONENT-1. Stop on the first success.
*> Mind that not all platforms might implement subnormal numbers.
*> (2) x = 2**m, where m = MINEXPONENT, ..., 0. Stop on the first success.
*> (3) x = OV, where OV is the overflow threshold. OV^2 overflows but the norm is OV.
*> (4) x = 2**m, where m = MAXEXPONENT-1, ..., 1. Stop on the first success.
*>
*> Tests:
*> (a) y = x + 0 * I, |y| = x
*> (b) y = 0 + x * I, |y| = x
*> (c) y = (3/4)*x + x * I, |y| = (5/4)*x whenever (3/4)*x and (5/4)*x can be exactly stored
*> (d) y = (1/2)*x + (1/2)*x * I, |y| = (1/2)*x*sqrt(2) whenever (1/2)*x can be exactly stored
*>
*> Special cases:
*>
*> (i) Inf propagation
*> (1) y = Inf + 0 * I, |y| is Inf.
*> (2) y =-Inf + 0 * I, |y| is Inf.
*> (3) y = 0 + Inf * I, |y| is Inf.
*> (4) y = 0 - Inf * I, |y| is Inf.
*> (5) y = Inf + Inf * I, |y| is Inf.
*>
*> (n) NaN propagation
*> (1) y = NaN + 0 * I, |y| is NaN.
*> (2) y = 0 + NaN * I, |y| is NaN.
*> (3) y = NaN + NaN * I, |y| is NaN.
*>
*> \endverbatim
*
program zabs
integer N, i, nNaN, nInf, min, Max, m
parameter ( N = 4, nNaN = 3, nInf = 5 )
double precision X( N ), R, threeFourth, fiveFourth, answerC,
$ answerD, oneHalf, aInf, aNaN, relDiff, b,
$ eps, blueMin, blueMax, Xj, stepX(N), limX(N)
parameter ( threeFourth = 3.0d0 / 4,
$ fiveFourth = 5.0d0 / 4,
$ oneHalf = 1.0d0 / 2 )
double complex Y, cInf( nInf ), cNaN( nNaN )
intrinsic ABS, DBLE, RADIX, CEILING, TINY, DIGITS, SQRT,
$ MAXEXPONENT, MINEXPONENT, FLOOR, HUGE, DCMPLX,
$ EPSILON
*
min = MINEXPONENT(0.0d0)
Max = MAXEXPONENT(0.0d0)
m = DIGITS(0.0d0)
b = DBLE(RADIX(0.0d0))
eps = EPSILON(0.0d0)
blueMin = b**CEILING( (min - 1) * 0.5d0 )
blueMax = b**FLOOR( (Max - m + 1) * 0.5d0 )
*
X(1) = TINY(0.0d0) * b**( DBLE(1-m) )
X(2) = TINY(0.0d0)
X(3) = HUGE(0.0d0)
X(4) = b**( DBLE(Max-1) )
*
stepX(1) = 2.0
stepX(2) = 2.0
stepX(3) = 0.0
stepX(4) = 0.5
*
limX(1) = X(2)
limX(2) = 1.0
limX(3) = 0.0
limX(4) = 2.0
*
print *, '# X :=', X
print *, '# Blue min constant :=', blueMin
print *, '# Blue max constant :=', blueMax
*
Xj = X(1)
if( Xj .eq. 0.0d0 ) then
print *, "# Subnormal numbers treated as 0"
else
do 100 i = 1, N
Xj = X(i)
if( Xj .eq. 0.0d0 ) print *,
$ "# Subnormal numbers may be treated as 0"
100 continue
endif
*
aInf = X(3) * 2
cInf(1) = DCMPLX( aInf, 0.0d0 )
cInf(2) = DCMPLX(-aInf, 0.0d0 )
cInf(3) = DCMPLX( 0.0d0, aInf )
cInf(4) = DCMPLX( 0.0d0,-aInf )
cInf(5) = DCMPLX( aInf, aInf )
*
aNaN = aInf / aInf
cNaN(1) = DCMPLX( aNaN, 0.0d0 )
cNaN(2) = DCMPLX( 0.0d0, aNaN )
cNaN(3) = DCMPLX( aNaN, aNaN )
*
* Test (a) y = x + 0 * I, |y| = x
do 10 i = 1, N
Xj = X(i)
if( Xj .eq. 0.0d0 ) then
print *, "# [a] Subnormal numbers may be treated as 0"
else
do while( Xj .ne. limX(i) )
Y = DCMPLX( Xj, 0.0d0 )
R = ABS( Y )
if( R .ne. Xj ) then
relDiff = ABS(R-Xj)/Xj
WRITE( *, FMT = 9999 ) 'a',i, Y, R, Xj, relDiff
endif
Xj = Xj * stepX(i)
end do
endif
10 continue
*
* Test (b) y = 0 + x * I, |y| = x
do 20 i = 1, N
Xj = X(i)
if( Xj .eq. 0.0d0 ) then
print *, "# [b] Subnormal numbers may be treated as 0"
else
do while( Xj .ne. limX(i) )
Y = DCMPLX( 0.0d0, Xj )
R = ABS( Y )
if( R .ne. Xj ) then
relDiff = ABS(R-Xj)/Xj
WRITE( *, FMT = 9999 ) 'b',i, Y, R, Xj, relDiff
endif
Xj = Xj * stepX(i)
end do
endif
20 continue
*
* Test (c) y = (3/4)*x + x * I, |y| = (5/4)*x
do 30 i = 1, N
if( i .eq. 3 ) go to 30
if( i .eq. 1 ) then
Xj = 4*X(i)
else
Xj = X(i)
endif
if( Xj .eq. 0.0d0 ) then
print *, "# [c] Subnormal numbers may be treated as 0"
else
do while( Xj .ne. limX(i) )
answerC = fiveFourth * Xj
Y = DCMPLX( threeFourth * Xj, Xj )
R = ABS( Y )
if( R .ne. answerC ) then
relDiff = ABS(R-answerC)/answerC
WRITE( *, FMT = 9999 ) 'c',i, Y, R, answerC, relDiff
endif
Xj = Xj * stepX(i)
end do
endif
30 continue
*
* Test (d) y = (1/2)*x + (1/2)*x * I, |y| = (1/2)*x*sqrt(2)
do 40 i = 1, N
if( i .eq. 1 ) then
Xj = 2*X(i)
else
Xj = X(i)
endif
if( Xj .eq. 0.0d0 ) then
print *, "# [d] Subnormal numbers may be treated as 0"
else
do while( Xj .ne. limX(i) )
answerD = (oneHalf * Xj) * SQRT(2.0d0)
if( answerD .eq. 0.0d0 ) then
print *, "# [d] Subnormal numbers may be treated as 0"
else
Y = DCMPLX( oneHalf * Xj, oneHalf * Xj )
R = ABS( Y )
relDiff = ABS(R-answerD)/answerD
if( relDiff .ge. (0.5*eps) )
$ WRITE( *, FMT = 9999 ) 'd',i, Y, R, answerD, relDiff
endif
Xj = Xj * stepX(i)
end do
endif
40 continue
*
* Test (e) Infs
do 50 i = 1, nInf
Y = cInf(i)
R = ABS( Y )
if( .not.(R .gt. HUGE(0.0d0)) ) then
WRITE( *, FMT = 9999 ) 'i',i, Y, R, aInf, aInf
endif
50 continue
*
* Test (f) NaNs
do 60 i = 1, nNaN
Y = cNaN(i)
R = ABS( Y )
if( R .eq. R ) then
WRITE( *, FMT = 9998 ) 'n',i, Y, R
endif
60 continue
*
9998 FORMAT( '[',A1,I1, '] ABS( ', (ES10.3,SP,ES10.3,"*I"), ' ) = ',
$ ES10.3, ' differs from NaN' )
*
9999 FORMAT( '[',A1,I1, '] ABS( ', (ES10.3,SP,ES10.3,"*I"), ' ) = ',
$ ES10.3,' differs from ', ES10.3, ' #relative diff = ', ES10.3 )
*
* End of zabs
*
END
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment