Re: [eigen] sse asin implementation

[ Thread Index | Date Index | More Archives ]

Hash: SHA256

And everyone who knows numerics is using only floats and the extended
precision of doubles only where it gives an advantage (e.g. to iterate
the error away).

Any algorithm that won't work with floats is badly conditioned and
probably won't work with doubles anyway...

Rohit Garg schrieb:
> Actually AFAIK, using more than 64 bits for doubles is _non_ compliant
> with the IEEE spec.
> On Tue, Apr 7, 2009 at 7:25 PM, Benoit Jacob <jacob.benoit.1@xxxxxxxxx> wrote:
>> 2009/4/7 Rohit Garg <rpg.314@xxxxxxxxx>:
>>> Hang on, I thought otherwise. And Gael pointed out that the sqrt_ps
>>> was exact according to IEEE spec, what about add,sub,mul,div?
>> Ah ok, i got that one wrong. A quick google search suggests that SSE
>> is IEEE compliant indeed. What remains true is that SSE is less
>> precise than non-SSE code for double's because in non-SSE the CPU is
>> able to use long double's of 80 bit internally to limit propagation of
>> errors, but I don't know if anything like that applies for floats
>> (since here we're discussing floats).
>> I told you i'm not competent ;)
>> Benoit

Version: GnuPG v1.4.9 (GNU/Linux)


Mail converted by MHonArc 2.6.19+