topic This is the usual issue with in IntelĀ® Fortran Compiler
https://community.intel.com/t5/Intel-Fortran-Compiler/new-versions-of-svml-give-different-results-that-older-versions/m-p/1143969#M137989
<P>This is the usual issue with floating point numbers. For floating point numbers you should never test for exact equality</P><P>but for equality within a given precision. E.g. something like the code below. Check the documentation on the flags for the numerical models in order</P><P>to look what the exact behavior is.</P>
<PRE class="brush:fortran; class-name:dark;"> elemental function nearly_equal_real (a, b, abs_smallness, rel_smallness) result (r)
logical :: r
real(default), intent(in) :: a, b
real(default), intent(in), optional :: abs_smallness, rel_smallness
real(default) :: abs_a, abs_b, diff, abs_small, rel_small
abs_a = abs (a)
abs_b = abs (b)
diff = abs (a - b)
! shortcut, handles infinities and nans
if (a == b) then
r = .true.
return
else if (ieee_is_nan (a) .or. ieee_is_nan (b) .or. ieee_is_nan (diff)) then
r = .false.
return
end if
abs_small = tiny_13; if (present (abs_smallness)) abs_small = abs_smallness
rel_small = tiny_10; if (present (rel_smallness)) rel_small = rel_smallness
if (abs_a < abs_small .and. abs_b < abs_small) then
r = diff < abs_small
else
r = diff / max (abs_a, abs_b) < rel_small
end if
end function nearly_equal_real</PRE>
<P> </P>
<P> </P>Sat, 11 May 2019 12:41:05 GMTJuergen_R_2019-05-11T12:41:05Znew versions of svml give different results that older versions for many math functions
https://community.intel.com/t5/Intel-Fortran-Compiler/new-versions-of-svml-give-different-results-that-older-versions/m-p/1143968#M137988
<P>Consider the following code as an example</P><P>version 14.0 of svml calculates the sin(90) as 1 while latest svml calculates it as 0.99999.</P><P>This can obviously create problems if subsequent calculations rely on the sin result to be exactly one.</P><P>Could someone please advice what is the best way to deal with these roundoff errors in newer versions of the libraries.</P>
<PRE class="brush:fortran; class-name:dark;">program sum
real angle_degrees(100)
real angle_radians(100)
pi = 4 * atan (1.0_4)
pi180 = pi/180.0
angle_degrees = 90
angle_radians = sin(angle_degrees * pi180 )
print * , angle_radians
end program sum
</PRE>
<P> </P>Sat, 11 May 2019 12:34:30 GMThttps://community.intel.com/t5/Intel-Fortran-Compiler/new-versions-of-svml-give-different-results-that-older-versions/m-p/1143968#M137988gn1642019-05-11T12:34:30ZThis is the usual issue with
https://community.intel.com/t5/Intel-Fortran-Compiler/new-versions-of-svml-give-different-results-that-older-versions/m-p/1143969#M137989
<P>This is the usual issue with floating point numbers. For floating point numbers you should never test for exact equality</P><P>but for equality within a given precision. E.g. something like the code below. Check the documentation on the flags for the numerical models in order</P><P>to look what the exact behavior is.</P>
<PRE class="brush:fortran; class-name:dark;"> elemental function nearly_equal_real (a, b, abs_smallness, rel_smallness) result (r)
logical :: r
real(default), intent(in) :: a, b
real(default), intent(in), optional :: abs_smallness, rel_smallness
real(default) :: abs_a, abs_b, diff, abs_small, rel_small
abs_a = abs (a)
abs_b = abs (b)
diff = abs (a - b)
! shortcut, handles infinities and nans
if (a == b) then
r = .true.
return
else if (ieee_is_nan (a) .or. ieee_is_nan (b) .or. ieee_is_nan (diff)) then
r = .false.
return
end if
abs_small = tiny_13; if (present (abs_smallness)) abs_small = abs_smallness
rel_small = tiny_10; if (present (rel_smallness)) rel_small = rel_smallness
if (abs_a < abs_small .and. abs_b < abs_small) then
r = diff < abs_small
else
r = diff / max (abs_a, abs_b) < rel_small
end if
end function nearly_equal_real</PRE>
<P> </P>
<P> </P>Sat, 11 May 2019 12:41:05 GMThttps://community.intel.com/t5/Intel-Fortran-Compiler/new-versions-of-svml-give-different-results-that-older-versions/m-p/1143969#M137989Juergen_R_2019-05-11T12:41:05ZJuergen,
https://community.intel.com/t5/Intel-Fortran-Compiler/new-versions-of-svml-give-different-results-that-older-versions/m-p/1143970#M137990
<P>Juergen,</P><P>Three points:</P><P>1) the epsilon function should be used in combination with the magnitudes of the input arguments such that the precision takes into account of these magnitudes.</P><P>2) ieee_is_... intrinsic functions have significant overhead, if possible perform what you can before making these tests.</P><P>3) While one NAN is not equal to one not-NAN, two NAN's are neither equal nor unequal</P><P>Jim Dempsey</P>Sat, 11 May 2019 13:18:16 GMThttps://community.intel.com/t5/Intel-Fortran-Compiler/new-versions-of-svml-give-different-results-that-older-versions/m-p/1143970#M137990jimdempseyatthecove2019-05-11T13:18:16ZThanks, Jim. Our definition
https://community.intel.com/t5/Intel-Fortran-Compiler/new-versions-of-svml-give-different-results-that-older-versions/m-p/1143971#M137991
Thanks, Jim. Our definition of "tiny" contains the epsilon function:
real(default), parameter, public :: &
eps0 = epsilon (zero), &
tiny_13 = 1E3_default * epsilon (zero), &
tiny_10 = 1E6_default * epsilon (zero), &
tiny_07 = 1E9_default * epsilon (zero)
Furthermore, we do test runs with signalling NaNs for our software, so that the test for NaNs in this float comparison function is rather academic.
But most definitely thanks for the remarks.Sat, 11 May 2019 13:22:38 GMThttps://community.intel.com/t5/Intel-Fortran-Compiler/new-versions-of-svml-give-different-results-that-older-versions/m-p/1143971#M137991Juergen_R_2019-05-11T13:22:38ZOops, strike 1), the r = diff
https://community.intel.com/t5/Intel-Fortran-Compiler/new-versions-of-svml-give-different-results-that-older-versions/m-p/1143972#M137992
<P>Oops, strike 1), the r = diff... should be sufficient to take into consideration the magnitudes verses precision.</P><P>Jim Dempsey</P>Sat, 11 May 2019 15:29:53 GMThttps://community.intel.com/t5/Intel-Fortran-Compiler/new-versions-of-svml-give-different-results-that-older-versions/m-p/1143972#M137992jimdempseyatthecove2019-05-11T15:29:53ZBut then these are relative
https://community.intel.com/t5/Intel-Fortran-Compiler/new-versions-of-svml-give-different-results-that-older-versions/m-p/1143973#M137993
<P>But then these are relative to 1E3, 6 and 9. When you have numbers of much larger (e.g. astronomical scale) or smaller (atomic scale) then the magnitude of your tiny(s) must be futzed with. You might want to consider</P><P>tiny = min(a,b) * epsilon(zero) * YouPickNumberRelativeToLSB (e.g 2, 4, ...)</P><P>Jim Dempsey</P>Sat, 11 May 2019 15:35:18 GMThttps://community.intel.com/t5/Intel-Fortran-Compiler/new-versions-of-svml-give-different-results-that-older-versions/m-p/1143973#M137993jimdempseyatthecove2019-05-11T15:35:18Z