On Wednesday 17 September 2003 21:26, Aleksandar Donev wrote:
> Hello,
>
> I am observing something I have not seen before: a massive loss of
> precision with one compiler (out of 3). I have some calculations which have
> certain mathematical properties, such as the norm of a quaternion being
> preserved at 1. With two compilers this is correct within numerical
> (double) precision, 1E-15 or alike. But with one compiler I am getting much
> larger discrepancies, order 1E-13 to 1E-12, and these eventually lead to my
> whole simulation breaking down. The calculations involved are many small
> matrix operations (inverses, multiplications), so it is quite imaginable
> there would be significant roundoff.
>
> What could this be related to? Is it optimization, possible mismatch
> between compiler and platform (the compiler is older then the processor),
> or should I be checking my codes?
A compiler for a CPU with extra precision operations could gain that much
precision by optimizing inner loops. A little more information could save us
from hopelessly wild guesses.
--
Tim Prince
|