Re: [eigen] (General question) Floating point: why are 'inf' and 'nan' slow?

[ Thread Index | Date Index | More lists.tuxfamily.org/eigen Archives ]


2009/9/23 Rohit Garg <rpg.314@xxxxxxxxx>:
> On Wed, Sep 23, 2009 at 10:27 PM, Benoit Jacob <jacob.benoit.1@xxxxxxxxx> wrote:
>> The list of x86 CPUs that don't have SSE2 (SSE is not enough for
>> double) includes Pentium 3, Athlon XP, VIA C7, AMD Geode, etc. There's
>> no way that we could neglect all of them performance-wise. Moreover,
>> even with SSE2, people may still want to use -mfpmath=387 (and it's
>> the default) in which case the non-vectorized part of Eigen
>> computations is affected.
>>
>> It's not a corner case at all, I was wondering if when redesigning the
>> solvers I could assume it to be safe to produce INF and NAN during the
>> computation, not just as return values at the end, and the answer is
>> that I can't do that.
>
> How do you propose to handle this then? Will you check for values
> before writing them to memory? And what happens on the SSE2 machines?
> There too, denormals are handled in sw.

I don't need to handle all cases, it's OK if in corner cases I have INF values.

All I need to take care of is the "mainstream" cases. This is far
easier :) Current Eigen already does that, as does LAPACK.

Benoit



Mail converted by MHonArc 2.6.19+ http://listengine.tuxfamily.org/