[gmx-developers] Verlet-Tables integration
erik.lindahl at gmail.com
Sat Nov 17 11:33:07 CET 2018
[Posting at the end of the thread to avoid divergence]
First, the memory-vs-flop question is mostly relevant when doing more of
one will save the other ;-) If we can avoid pushing e.g. 2*4*4=32 bytes of
extra table memory data by doing SIMD computation (e.g. as we do for
analytical Ewald), that can be a clear win. However, adding more
floating-point operations in our very simplest interaction without removing
memory ops is not going to be a clear win ;-)
Second, we will still need 1/sqrt(x) and/or 1/x for the normal
interactions, so I think it would lead to more complex kernels that need to
decide what to compute in each case. Nothing is impossible, and the new
templated kernels we are planning might make it a bit easier for people to
do - but it's an added complication for everyone hacking the kernels that
I'd prefer to avoid until we know it would broadly used.
A long time ago we thought about tabulating the scaled interaction instead
(to combine dispersion/repulsion), but that does not work with switched
interactions (since the switching range would also be scaled). Combining
"everything" into one table is a fun idea, but for OPLS I would assume that
would lead to some ~400 different tables (normally there are about 30
distinct atom types remaining after pruning), which means there won't be
enough data in most inner kernel iterations.
Finally, while all-table interactions can be a bit slow, remember that we
virtually never need to tabulate van der Waals interactions in normal runs,
so it's not going to influence performance at all there.
On Sat, Nov 17, 2018 at 10:44 AM Berk Hess <hess at kth.se> wrote:
> But then we would need different coordinate and force buffers for each
> table pair, as well as a separate grid for each such buffer. Technically
> possible, but this will not be fast, unless we have only 2 or 3 tables.
> Gromacs Developers mailing list
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/GMX-developers_List before
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> * For (un)subscribe requests visit
> or send a mail to gmx-developers-request at gromacs.org.
Erik Lindahl <erik.lindahl at dbb.su.se>
Professor of Biophysics, Dept. Biochemistry & Biophysics, Stockholm
Science for Life Laboratory, Box 1031, 17121 Solna, Sweden
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the gromacs.org_gmx-developers