CPLAtof() - different results than atof()

[warmerdam]

The following precision different between atof() and CPLAtof() is having annoying "knock on" affects on generation of EPSG definition files.

warmerda@amd64[45]% cat test1.cpp 
#include "gdal.h"
#include "cpl_conv.h"

int main()

{
        double x1, x2;
        const char *str = "6356566.435";

        x1 = atof(str);
        x2 = CPLAtof(str);

        printf( "%.16g, %.16g\n", x1, x2 );
}
warmerda@amd64[46]% ./test1
6356566.435, 6356566.435000001
(gdb) run
Starting program: /wrk/home/warmerda/test1 
[Thread debugging using libthread_db enabled]
[New Thread 46912564802784 (LWP 354)]
[Switching to Thread 46912564802784 (LWP 354)]

Breakpoint 1, main () at test1.cpp:8
8               const char *str = "6356566.435";
(gdb) n
10              x1 = atof(str);
(gdb) 
11              x2 = CPLAtof(str);
(gdb) 
13              printf( "%.16g, %.16g\n", x1, x2 );
(gdb) p x1
$1 = 6356566.4349999996
(gdb) p x2
$2 = 6356566.4350000005

[dron]

I have reproduced this problem and also constructed few more cases when CPLAtof() results differs from the libc's atof(): some of my tests started to fail when I reduced the tolerance threshold to 0.0000000000000001. It seems that CPLAtof() typically overestimates the value, but the standard atof() underestimates it.

I will take a look what can be done about that (though, the whole floating point topic is a delicate matter).

By the way, maybe we should not use two different atof() implementations in the same software? It is pretty easy to add the locale aware variant of the CPLAtof() and stop using the standard atof() completely. Otherwise it is possible that we will constantly come to issues like that in the future.

Best regards, Andrey

[msieczka]

This also affect PROJ 4.6.0.

Compare the PROJ.4's epsg file in version 4.5.0 and 4.6.0, and see how parameters differ in many SRS definitions, eg. EPSG 26794-26798, 2180, 2255-2262, 2057, 2000-2007 and more. Eg.:

4.5.0:
<4009> +proj=longlat +a=6378450.047548896 +b=6356826.621488444 +no_defs <>

4.6.0:
<4009> +proj=longlat +a=6378450.047548897 +b=6356826.621488445 +no_defs  <>

4.5.0:
<2965> +proj=tmerc +lat_0=37.5 +lon_0=-85.66666666666667 +k=0.999967 +x_0=99999.99989839978 +y_0=249999.9998983998 +ellps=GRS80 +datum=NAD83 +to_meter=0.3048006096012192 +no_defs  <>

4.6.0:
<2965> +proj=tmerc +lat_0=37.5 +lon_0=-85.66666666666667 +k=0.999966667 +x_0=99999.99989839978 +y_0=249999.9998983998 +ellps=GRS80 +datum=NAD83 +to_meter=0.3048006096012192 +no_defs  <>

4.5.0:
<2180> +proj=tmerc +lat_0=0 +lon_0=19 +k=0.999300 +x_0=500000 +y_0=-5300000 +ellps=GRS80 +units=m +no_defs  <>

4.6.0:
<2180> +proj=tmerc +lat_0=0 +lon_0=19 +k=0.9993000000000001 +x_0=500000 +y_0=-5300000 +ellps=GRS80 +units=m +no_defs  <>

[dron]

The problem is in the way how _Ldmul() function works. To fix it that function (or, more generally, _Ldtento()) should be reimplemented. I will think what can I do about that.

[dron]

This is should be fixed now with the r14379.

[msieczka]

Please backport this fix to stable branch, as it's Milestone is 1.5.2.

[mloskot]

I'm taking the porting task on myself

[rouault]

Mateusz,

I'd suggest you to wait a bit about #2036 as there's a pending issue with the fact that CPLStrtof and CPLStrtofDelim have been removed. See  http://lists.osgeo.org/pipermail/gdal-dev/2008-May/017031.html

[mloskot]

Even,

You are right. I'll wait until this issue is clarified.

[dron]

Folks, I have restored these two functions back. See r14464.

[msieczka]

Hi Mateusz

Do you think it would be now possible for you to backport the fix into the stable branch?

Cheers, Maciek

[mloskot]

Ported recent fixes to branches/1.5 (r14568). Closing.