CGAL users discussion list

[Ben Supnik <>]

Hi Hichem,

BBB HHH wrote:
> I think the problem is not related to the -frounding-math version 
> because of the successful execution of the second version of source 
> code. Moreover, i am using exact number types.

Wait - I think this does not follow.

IF the CGAL code is correct:

- If rounding math is correct, boolean ops will succeed 100% of the time.
- If ronding math is incorrect, boolean ops will fail < 100% of the time.

In other words, the absence of evidence is not the evidence of absence. 
 I went a few weeks before I realized my rounding math was 
off...because a lot of the time the code succeeds.

Rounding math affects the exact kernel based on MP_float and quotient. 
I do not know how it affects gmp-based kernels or any of the 
filters/wrappers you might put around a kernel.

(My guess is that the lazy-exact-NT wrapper around your kernel might 
depend on rounding math for it's interval calcs.)

I am not saying here that you have a rounding problem...just that those 
two conditions (successful execution of the same data with other code 
and the use of an exact kernel) exactly match the case where I _did_ 
have a rounding error.

But...that may be moot -- we will see what Efi says.

Actually one thing to try: remove the lazy_exact wrapper from your 
kernel - just use simple_cartesian<gmpq>.  Besides being really slow, 
does the first code work? :-)

cheers
Ben