[Numpy-discussion] A weekend floating point/compiler question
Fernando Perez
Fernando.Perez at colorado.edu
Fri Apr 28 16:55:02 EDT 2006
Hi Robert and George,
We found a bug in g77 v. 3.4.4 as well as in gcc, which manifests itself in
the following little snippet:
planck[f77bug]> cat testbug.f
program testbug
c
implicit real *8 (a-h,o-z)
c
half = 0.5d0
x = 0.49d0
nnx = 100
iax = (x+half)*nnx
print *, 'Should be 99:',iax
stop
end
c EOF
planck[f77bug]> g77 -o testbug.g77 testbug.f
planck[f77bug]> ./testbug.g77
Should be 99: 98
This can be seen as computing (x/n+1/2)*n and comparing it to x+n/2. Greg is
using this in a number of places inside a library, which had never given
trouble before when built with other compilers, like the sun, IBM, Intel and
Lahey ones. Now with g77 it gives the result above.
Questions:
1. Have you seen similar behavior in the past?
2. If we switch away from g77, what do you suggest moving towards? We ran
paranoia on ifort, lahey and g77, and lahey was the best performing of all.
The intel one has the advantage of being free. On the other hand, paranoia
did complain about arithmetic issues with it (though the above code works fine
with intel).
Any ideas you can give us would be very appreciated.
Cheers,
Fernando and Greg.
ps. Apparently g77 v 3.3.2 does NOT have this problem.
More information about the NumPy-Discussion
mailing list