"make check" failure on s390x architecture

[Andy Moorhouse]

Hello!

I got an error from "make check" in the wkml tests on our platform (which is probably a bit obscure: Red Hat Enterprise Linux 6.7 running on an IBM z13 mainframe under z/VM virtualisation - architecture is "s390x").

The error is a slight rounding issue:

./test_kmlContours_3_real
------------
--- test.xml    2016-08-10 12:58:51.053957837 +0000
+++ ./test_kmlContours_3_real.xml       2012-01-05 10:04:10.000000000 +0000
@@ -134037,7 +134037,7 @@
                     -2.534693,53.973289
                     -2.534689,53.974190
                     -2.534082,53.975101
-                    -2.533841,53.976002
+                    -2.533842,53.976002
                     -2.534302,53.976908
                     -2.534930,53.977599
                     -2.535281,53.977809
------------

I will ignore this as it is such a small rounding error, but thought I should put it "on the record", and I'm interested to see if there will be any responses!

Many thanks
Andy

[Andrew Walker]

Hi Andy,

Thanks for the report. This is the only part of FoX with any significant floating point arithmetic so I suppose seeing a rounding issue here makes sense. It may also be relevant that this is all done with Fortran default reals. I don't think there is any guarantee that these are always the same precision on different systems, but I've never noticed a difference in practice and don't know what is used on your platform. We are probably using one or two too many decimal places in the tests but as we never got around to creating a double precision version of the contouring code it's not easy to be sure.

I assume this is with the gcc fortran compiler?

Best wishes,

Andrew

[Andy Moorhouse]

Hello Andrew, great to see such a quick response, thanks!

Yes, apologies, I did forget to mention that was with gfortran: gcc version 4.4.7 20120313 (Red Hat 4.4.7-16) (GCC)

On this platform we also have a pretty ancient NAG compiler, which NAG built quite some while ago for us on this platform: NAGWare Fortran 95 compiler Release 5.1

So I thought I'd try the build against the NAG compiler, to see if it was an issue with the underlying platform or the compiler. Interestingly, I again got a single error from the wkml checks, but it was not the same one as with gfortran!

cat failed.out
./test_kmlCreatePointStyle_7
------------
--- test.out    2016-08-15 07:36:46.674048076 +0000
+++ ./test_kmlCreatePointStyle_7.out    2012-01-05 10:04:10.000000000 +0000
@@ -0,0 +1,4 @@
+ Z90000FF
+           1
+ Invalid color value
+           0
------------

We will be using gfortran and not NAG, so I am not concerned with this at all. But it seems to show that the original rounding issue is with gfortran and not the platform, which might be interesting to others.

Regards and thanks for your help (and for Fox itself, which I believe will be extremely useful to our organisation).
Andy