[gmx-developers] Re: Cubic splines - again
lindahl at cbr.su.se
Tue Aug 22 15:00:01 CEST 2006
On Aug 22, 2006, at 12:46 PM, Berk Hess wrote:
> But I think we would still prefer to have continuous second
> derivatives as well.
> So the problem is then not in the interloops, but in the generation
> of the tables.
> There we should not use the second derivative of the potential,
> but rather the second derivative of the cubic spline fit of the
Agreed. In principle, the current innerloop code will work unmodified
for any third-power splines, including e.g. akima.
Further, on most architectures the costly step is to (a) calculate a
table index (truncation, etc.) and (b) fetch the table data from
memory. The amount of floating-point calculations involved is not
critical (though not entirely unimportant).
If we come up with a better table implementation this part is _very_
easy to change, including the assembly implementations. Not quite
search-and-replace, but almost.
More information about the gromacs.org_gmx-developers