CGAL users discussion list

[Marc Glisse <>]

On Thu, 28 Feb 2013, marcosscriven wrote:

> The thing is we *can* call into C/C++ code, it's just we have to implement
> it ourselves (so that we can compile it into Javascript too).

Well, no, I meant call into *natively compiled* C code ;-)

> The only implementations I can find rely on the CPU to do the right 
> rounding, I've not found a code implementation of one.

Implementing fesetround can hardly be anything but a use of the 
CPU-specific set-the-rounding-mode instruction, so if all you can program 
in is core C there is no way around that.

> I.e, change the current mode of rounding in the processor itself,
> temporarily, just for certain functions?

Yes.

> Is it right to say therefore it's not possible to choose just one rounding
> mode, and use that the whole time?

If you use only static filters and no intervals, you might get away with a 
constant round-to-nearest mode. If you use intervals (and not all 
operations on them, no sqrt in particular) but no static filters, you 
might get away with a constant round-upwards mode. If you were extremely 
motivated, you could recompute the bounds in the static filters to make 
them loose enough to work with any rounding.

> Is it also correct to say that the reason there would be infinite loops is
> that a numerical method relying on a particular rounding mode might never
> meet its convergence condition?

The reason is that predicates would give wrong answers. And as explained 
in length on the CGAL website, this may yield combinatorially impossible 
configurations, obviously unplanned for in the algorithms, and thus you 
might for instance walk in circles in a triangulation trying to locate 
a point.

> If so, is there any way in CGAL to specify an 'acceptable' accuracy?

No. The exact computation paradigm is not about getting "close enough" 
results, it is about remaining consistant to ensure we get a result at 
all. Proper rounding is essential in determining if the approximate 
computations were acceptably accurate or if we need to compute things 
exactly (with gmp).

> Thanks again for your help - appreciate this is an unusual situation.

Best solution might be to introduce fe[sg]etround functions in your 
favorite javascript engine.

--
Marc Glisse