[Thread Prev][Thread Next][Index]

Re: [ferret_users] differences in conversion of real*4 constants

Hi Torsten,
Thanks for the report and the example script. What you're seeing is entirely due to how numbers are evaluated with the grave-accent syntax. Grave accents are evaluated as the command is read in, at the command-parsing step. So for instance in the LIST command below, the /PREC=7 has no effect. When running interactively, you'll see the command echoed back with grave-accent expressions evaluated, and that is exactly the command that is run by Ferret.  So `xa` is always evaluated as -5.0008 and so on.  There's a setting that has these numbers evaluated only to 5 digits, less than full single precision.
yes? ! def_0: define axes with predefined constants
yes? let nx=577; let ny=257;
yes? let xa=-5.00081; let xe=31.2892
yes? let ya=51.8717;  let ye=68.0006
yes? def axis/x=`xa`:`xe`/npoints=`nx`/units="degrees_east"  xax0
 !-> def axis/x=-5.0008:31.289/npoints=577/units="degrees_east"  xax0
yes? def axis/y=`ya`:`ye`/npoints=`ny`/units="degrees_north" yax0
 !-> def axis/y=51.872:68.001/npoints=257/units="degrees_north" yax0

yes? list/prec=7 `xa`,`xe`,`ya`,`ye`
 !-> list/prec=7 -5.0008,31.289,51.872,68.001
 Column  1: cnst is constant
 Column  2: cnst is constant
 Column  3: cnst is constant
 Column  4: cnst is constant
              cnst     cnst      cnst      cnst
I / *:    -5.000800  31.28900  51.87200  68.00100

What you would want is this (shortened listing)
yes? list/nohead/prec=7 `xa`,`xe`,`ya`,`ye`
 !-> list/nohead/prec=7 -5.00081,31.2892,51.8717,68.00060
I / *:    -5.000810  31.28920  51.87170  68.00060

Yours seems to me to be a reasonable request - one should be able to get back full single precision when evaluating a single precision number.  I've put this on our list to change in the next release.


Torsten Seifert wrote:
Dear Ferret users

please try the attached script, which demonstrates that the conversion of float constants may give different results (here using explicit constants an/or the `xa` syntax in definitions of axes).
The deviations are small, but definitely larger than is to be expected for real*4 (with a 24 bit mantissa the 7th digit may change, but not the 6th digit).
I am wondering how that could happen. It is clear, that a constant assigned to a variable by let is converted into real*4 accuracy. I would assume that the same conversion is done if the constants are specified explicitly in the axis definition because they are assinged to internal buffers.
Listing the same constants with Fortran shows, that def_1 is right since no changes appear up to the 7th digit. Thus the conversions by `xa` etc. seem to give erroneous results. Couls that be fixed, please?

Best regards


[Thread Prev][Thread Next][Index]

Contact Us
Dept of Commerce / NOAA / OAR / PMEL / TMAP

Privacy Policy | Disclaimer | Accessibility Statement